# The Agents Were Fine. Everything Around Them Broke. - Joey Wang, Harvey Joey Wang, Harvey [Talk page](https://ai.engineer/talks/ktbAVYsAxpk-agents-were-fine-everything-around-them-broke) · [Video](https://www.youtube.com/watch?v=ktbAVYsAxpk) ## The Agents Were Fine. Everything Around Them Broke. After 10,000 agent-written pull requests at Harvey, Joey Wang would build the CI follow-up loop first. He’d also put limits on what could start a task. One release bug sent about 900 agents after the same fix. ### What a cloud agent needs from Harvey [0:00](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=0s) In the month before Joey Wang’s talk, Spectre, Harvey’s internal cloud agent platform, started more than one agent per minute on average, day and night. About 460 people used it that month. Agents had written 10,000 pull requests, and people were finding other uses for the same setup. They asked data questions in Slack and got answers back. Someone in marketing dropped a prototype and a talk track into a channel and asked for help turning them into a story for a product video. Why build Spectre when engineers already have coding agents they like? Spectre runs those agents. Wang put the practical limit on a laptop at roughly five to ten at a time. Moving work to the cloud buys capacity, but a remote agent still needs a repository checkout, credentials and access to Harvey’s stack. It needs a way to pick up an alert or ticket and get a change through review. Spectre supplies that harness. Sharing it across the company adds another responsibility. Even when one engineer has plenty of work to hand off, the next person needs room to run. ### A 2 a.m. fix still has to pass CI [2:38](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=158s) Engineers adopting Cursor had two immediate questions. How could they get access to Harvey’s secrets, and how could they test against the real stack? Wiring both into Cursor would leave the team with the same job the next time engineers switched tools. Harvey let engineers ask their preferred agent to use Spectre. A session could start on a laptop, move to the cloud and come back. The team could keep its test environment as engineers changed agents. Take Wang’s example of an alert arriving in Slack at 2 a.m. An agent picks it up without anyone typing a prompt, writes a fix, spins up a copy of Harvey’s stack and posts evidence on the pull request. If continuous integration (CI) fails, the failure itself becomes input to the next attempt. A follow-up loop reads what failed and pushes another fix while the owner is asleep. By morning the PR is in the shared review queue under the same rules as any other change. Spectre waits until the owner’s working day begins to send a nudge. Once the change is approved, an agent can merge it automatically. The follow-up loop was the stage Wang would build first. Harvey enabled it for everyone in early October. Of the agent-written PRs that merged into Harvey’s backend that month, four out of five had gone through the loop. Testing a tricky service can depend on something one engineer knows, right down to which flag to leave alone. Spectre gave that knowledge a home in projects. People could record what mattered, agents could record what they learned, and a team could pin the skills it trusted. Any agent working in the project could draw on that knowledge. The project also kept a record of who ran what. Wang’s view was that whichever agent an engineer picked, the checks and the audit trail should be the same. ### The command line lets one agent put others to work [5:19](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=319s) In May, people usually opened a browser tab or tagged Spectre in Slack. Exactly one task that month came from a terminal. The team adopted a rule that any action a person could click should also be callable by an agent. The command-line interface could start a task, turn on the follow-up loop, enable auto-merge or publish a page. An agent on a laptop, an agent already inside a sandbox and a CI job could all use those commands. Where did the new work come from? By October the command line was starting about 6,500 tasks a month, while web-app usage held up. Some people started hundreds of tasks a month without ever opening the app. One conversation with an agent could now set a whole batch of work in motion. Wang used that access in two directions over the course of October. - **Hand work out.** Wang asked a coding session to coordinate a large ticket. It divided the work into ten milestones, kept one for Wang and handed the rest to Spectre as hundreds of tasks. Each task got its own sandbox for end-to-end validation. - **Bring decisions back.** Later that month, another agent watched Spectre for questions that were holding up work. It listed the questions, drafted Wang’s replies and sent them back once Wang approved them. Wang could answer workers without visiting each task to find out what it needed. Each sandbox also needed code to work on. When the team measured where task time went, the slowest step was Git checkout. Checking out all of Harvey’s repositories took more than a minute on average. The team kept a copy of every repository on every machine and let each task borrow the local copy. A burst of tasks on the same machine could reuse repositories the machine already had. ### One bug started about 900 investigations [7:38](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=458s) A mid-August release shipped a bug that fired an alert once per workspace. Hundreds of alerts landed in one Slack channel. A workflow tagged the agent on every alert, and every tag started a task. On an ordinary day that workflow started about 50 tasks. During the incident it started around 900 in four hours, all chasing the same bug until the cluster filled up. What should happen when the next workspace reports the bug? Under the original workflow, that report was an instruction to launch a worker. With the fix, a repeat alert joined the investigation already open. The same investigation could collect reports from many workspaces. Where the next workspace alert goes Follow the same repeat alert through the two versions of the workflow. Its destination determines whether another agent starts. - Another workspace reports the bug — Its alert lands in the shared Slack channel. - Tag the agent again — The workflow tags the agent on every alert. - Launch another task — Another agent starts investigating the same bug. - Join the open investigation — The investigation already under way receives the repeat alert. - Another workspace reports the bug → Tag the agent again (Every alert triggers a tag) - Tag the agent again → Launch another task (Each tag starts a task) - Another workspace reports the bug → Join the open investigation (Route the report to existing work) ### A fair launch queue for a shared cluster [8:25](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=505s) Two months later, an engineer did exactly what the team had encouraged and handed work to Spectre. One laptop submitted about 950 command-line tasks in about three hours. The cluster filled, and everyone else’s work sat behind that laptop. Routing repeat alerts to an open investigation addressed the August failure. This time Spectre needed a way to accept a large batch and still let other people work. Four days later the team had a fair launch queue. Submitting a task gave its owner a visible place in line. A dispatcher started work when the cluster had room, giving priority to people waiting at a keyboard. Each user had a fair share of capacity and could borrow more while the cluster was quiet. A large batch could use idle machines while the dispatcher retained control over who got the next opening. What happened to completion as volume rose? On the busy August day Wang used for comparison, about one task in ten failed to finish. On November 4, Spectre handled more than twice that day’s volume, and a larger share of tasks finished. Several changes landed together, so Wang couldn’t say which fix contributed most. ### The merge queue became the next bottleneck [10:06](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=606s) At the end of October, Harvey’s developer platform team raised a different alarm. Company-wide PR volume had jumped around 60% in one month. Merging a change used to take 15 to 30 minutes. Now it took over an hour, and the team repeatedly had to switch off the merge queue. With the queue off, a PR rebased onto `main`, reran CI and merged if the checks passed. Every PR depended on that shared starting point. If `main` broke, everyone’s CI broke with it. The fair launch queue could make room for people to start work. How to keep their finished changes moving through merge was still open. ### Who starts the work when nobody is watching? [10:45](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=645s) Spectre’s fastest-growing source of work was automation. A schedule or an event could wake an agent to triage an alert, read a PR or write a post. About 110 people had automations running in the month before the talk. An unattended agent could start thousands of tasks, so each automation switched itself off after five consecutive failures. Its owner could also give it a budget or an expiry date. Wang wondered what would happen if every Slack channel had a persistent agent that could follow a conversation and recognize when to begin work on its own. For reviewers, the question was whether every PR could arrive with evidence that the change worked before a human even looked at it. At the time of the talk, only one merge in 20 carried that kind of proof. And even a change that came with proof would face the question Harvey had yet to answer. How do you merge when agents keep opening PRs? ## Key takeaways - Give the platform responsibility for what follows the first patch. In Spectre, a failed CI check could drive another attempt before the reviewer’s working day began. - Make each user action callable. A command for starting a task or enabling follow-up gives a laptop agent, a sandboxed worker and CI access to the same capability. - Set limits before a person, agent or alert discovers how much work it can start. Routing repeat alerts to the open investigation prevents duplicate work. A fair dispatcher handles the different case where one person has hundreds of tasks to run. ## Slides ### 10,000 agent PRs later, what broke wasn’t the agents After 10,000 agent-written pull requests at Harvey, Joey Wang calls the agents “the easy part.” The talk follows what broke around them as they worked across the company’s repositories. [0:14](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=14s) [Title slide for Joey Wang of Harvey at AIE Code Summit SF 2026. The subtitle asks what breaks when coding agents touch every repository you own.](https://assets.ai.engineer/sites/aiecode2025/sha256/b2ba9f45d26a0f2d2741a602f168302021a0c29f3638e5f176292d771dfc0006.webp) AIE CODE SUMMIT SF 2026 · JOEY WANG, HARVEY 10,000 agent PRs later, what broke wasn’t the agents What breaks when coding agents touch every repo you own ### More than one task a minute Spectre handled about 10,000 tasks in May and about 45,000 in October, when roughly 460 people at Harvey used it. Across its seven app repos, Harvey had merged 10,000+ agent-written PRs. [0:34](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=34s) [Spectre reports 10,000+ agent-written pull requests merged across seven app repositories, about 45,000 tasks in October versus 10,000 in May, and about 460 people who used it in October.](https://assets.ai.engineer/sites/aiecode2025/sha256/e9a0871b01dbdade2fc3fb0766068add6f48afa26d018515018e730cd448299e.webp) SPECTRE TODAY More than one task a minute 10k+ agent-written PRs merged across our 7 app repos ~45k tasks in October, up from ~10k in May ~460 people used it in October ### Not just code Spectre went from about 10,000 tasks in May 2026 to about 45,000 in October. Code remained the largest category each month, even as teams used Spectre to answer data questions in Slack, triage incidents and storyboard marketing videos. [0:53](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=53s) [Monthly task totals from May through October 2026 are approximately 10k, 15k, 24k, 28k, 29k and 45k. Code is the largest segment each month. The legend also describes data and analytics, triage and monitoring, and creative work.](https://assets.ai.engineer/sites/aiecode2025/sha256/cca64cc1c55c3cf92113db867457782d2d0c43fc83328c36746593a651dfe032.webp) SPECTRE TODAY Not just code Tasks per month, by kind of work, 2026 May: ~10k Jun: ~15k Jul: ~24k Aug: ~28k Sep: ~29k Oct: ~45k Code Writes, reviews and fixes PRs Data and analytics Answers data questions in Slack Triage and monitoring Alerts, incidents, rollouts Creative Storyboards, animations and video graphics for marketing ### The harness, not the model When teams need to see what an agent did, retry failed tasks and track costs across providers, they rely on the software around the model. Harvey describes that shift in Spectre: “the harness becomes the product.” [1:15](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=75s) [A quotation from Harvey’s April 7, 2026 blog post “Building Spectre” lists durability, shared visibility, retries, cost accounting, multiple providers and collaboration as requirements. Its closing phrase, “the harness becomes the product,” is highlighted.](https://assets.ai.engineer/sites/aiecode2025/sha256/8501808452260436df0465ce12968d2b8a44e5abcf2ddd491fb0f944e515e1b1.webp) THE HARNESS, NOT THE MODEL “A thin wrapper is fine until you need durability, shared visibility, retry semantics, cost accounting, multi-provider support, and collaboration surfaces. At that point the harness becomes the product.” “Building Spectre,” Harvey blog, April 7, 2026 ### Why build, when Claude Code and Cursor exist? Harvey runs Claude Code, Codex and Cursor in its own sandboxes. The layer Harvey built around them gives each task a short-lived token and applies the same rules to every agent. That shared layer is what broke. [1:46](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=106s) [Five cards describe scoped tokens, validation, task triggers, shared policy and fair capacity. Beneath them, the slide explains that Claude Code, Codex and Cursor run inside Harvey’s sandboxes and that the surrounding layer Harvey built is what broke.](https://assets.ai.engineer/sites/aiecode2025/sha256/9af85356c5dd891c0f56bc5dfd8a2e32041d04fbe72dc0501a4bfb3d1336d5e6.webp) WHY BUILD Why build, when Claude Code and Cursor exist? We don’t replace them. We run them. Secrets in one place Scoped, short-lived tokens per task Our real stack Validation with our data and access Our triggers Alerts, Linear, schedules, PR events One policy Same rules and audit trail for every agent Fair capacity Hundreds of people, their bots and their coordinators Claude Code, Codex and Cursor run inside our sandboxes. We build the layer around them, and that layer is what broke. Interface text: So I want to tell three stories today. ### Three stories Harvey stopped competing on code generation and built a system that any coding agent could drive. The third story follows what broke in the infrastructure Harvey built around those agents. [2:28](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=148s) [Three numbered agenda items read “We stopped competing on writing code,” “We let any agent drive,” and “Then it broke.” A large black H fills the right panel.](https://assets.ai.engineer/sites/aiecode2025/sha256/7a8ba39c8e17d418cd9a11760889776a1287209182723761e497748916d2e69c.webp) AGENDA Three stories 1 We stopped competing on writing code 2 We let any agent drive 3 Then it broke H ### Pick your agent interface All six ways to start work lead to Spectre. Whichever interface an engineer chooses, the run has one record, one sandbox and the same later stages. The engineer can move the session from a laptop to the cloud and back. [2:43](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=163s) [Arrows from Claude Code, Codex, Cursor, ChatGPT Dots, Slack or web, and By hand point to Spectre. All six share one run record, one sandbox, and the same later stages. Notes describe skills and plugins, agents running inside the sandbox, and handing a session to the cloud and back.](https://assets.ai.engineer/sites/aiecode2025/sha256/35141c8b3061311e4388c6b7169c955f96ab0e1c97fbb86333011c01126655d0.webp) STORY 1 · AFTER THE CODE Pick your agent interface Claude Code Codex Cursor ChatGPT Dots Slack or web By hand Spectre underneath One run record, one sandbox, the same later stages Laptop agents call it through skills and plugins Claude Code, Codex or Cursor runs inside the sandbox Hand a session to the cloud and back That's how Spectre sits underneath all of it. ### Writing the code is step one Whichever agent writes the fix, it tests against a real stack and posts evidence on the PR. A follow-up loop fixes CI failures and keeps the PR current with main. Every PR enters one review queue. After approval, the agent arms auto-merge. [3:46](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=226s) [Six numbered cards describe event triggers, end-to-end validation with evidence on the pull request, a follow-up loop that fixes CI and keeps up with main, a shared review policy and queue, working-hours reminders, and agents arming auto-merge. The stages apply whichever agent wrote the code.](https://assets.ai.engineer/sites/aiecode2025/sha256/d3eb35baa621db36c08c2ff5b2f9b95427a05ff113c7dd9eef248bb76f094ab8.webp) STORY 1 · AFTER THE CODE Writing the code is step one 1 Triggers Slack, GitHub, Linear and schedules 2 E2E validation A real stack, evidence on the PR 3 Follow-up loop Fixes CI, keeps up with main 4 Review policy One policy and one queue for every PR 5 Nudges Reminds owners, in working hours 6 Merging Agents arm auto-merge, never merge Whichever agent wrote the code, the same stages run after it. ### 4 in 5 Of the agent-written PRs Harvey merged into its backend in October, 4 in 5 went through the follow-up loop. If a team can build only one stage after writing the code, Joey Wang recommends building that loop. [4:18](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=258s) [Four in five agent pull requests merged into the team’s backend in October went through the follow-up loop.](https://assets.ai.engineer/sites/aiecode2025/sha256/e8afb978034b3f6f63dfbdcecc22746fb31dae9f9743a62159a36a6d66f1d29a.webp) STORY 1 · AFTER THE CODE 4 in 5 agent PRs merged into our backend in October went through the follow-up loop ### One project holds what every agent needs Four agents share one Spectre project where people write the charter, agents record what they learn, and the team pins skills. Wang’s personal view is that the audit trail, readiness check and validation should be the same across agents. [4:52](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=292s) [Claude Code, Codex, Cursor and ChatGPT Dots point to one Spectre project with a charter, memory, skills, artifacts, role-based sharing and turn history. The speaker recommends the same audit trail, readiness check and validation across agents.](https://assets.ai.engineer/sites/aiecode2025/sha256/eae9fc1df8b126981064e22337386c6e59c89f2034d778fd36073a3c65bc4765.webp) STORY 1 · AFTER THE CODE One project holds what every agent needs A SPECTRE PROJECT Charter Written by people Memory Written by agents Skills Pinned for the team Artifacts Pages anyone can reuse Sharing People and teams, by role History Who ran each turn Claude Code Codex Cursor ChatGPT Dots MY PERSONAL VIEW, NOT HARVEY’S POSITION Whatever agent you use, the audit trail, readiness check and validation should be the same. ### If you can click it, an agent can call it Laptop and sandbox agents can use Spectre’s CLI to queue a task, keep a pull request current, arm auto-merge and publish a report. CI can use the same CLI to start or cancel tasks. [5:32](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=332s) [Terminal examples show task creation, pull-request follow-up, arming auto-merge and publishing a report. Callouts explain how laptop agents, sandboxes and CI can invoke Spectre.](https://assets.ai.engineer/sites/aiecode2025/sha256/54d6eb65077126b2e3a3989b62750200436ff301dbacd4bfdd515c205118aee4.webp) STORY 2 · ANY AGENT CAN DRIVE If you can click it, an agent can call it $ spectre task create "fix the flaky date test" ✓ Queued in a fresh sandbox $ spectre babysit ✓ Follow-up loop on: fixes CI, keeps up with main $ spectre pr auto-merge ✓ Auto-merge armed. The agent never merges. $ spectre artifact publish report.html --name "Q4 plan" ✓ Share link ready $ Laptop agents call it through skills and plugins Sandboxes call it too, with a scoped 4-hour token CI can start and cancel tasks Agents can drive Spectre, not just people. ### Tasks moved out of the web app Monthly command-line task starts rose from 1 in May 2026 to about 6,500 in October. The web app’s share of task starts fell from 40% to 19% as other ways of starting tasks grew faster. Web task volume held up. [5:59](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=359s) [Monthly command-line task starts rose from 1 in May 2026 to about 6,500 in October. The web app’s share fell from 40% to 19%. In October, about 1,700 pull requests merged from command-line tasks and 1,200 from web tasks.](https://assets.ai.engineer/sites/aiecode2025/sha256/f28f952ec0b51d501f0520665128ed477c9c4e64dff4ca07f253829ec7763ecc.webp) STORY 2 · ANY AGENT CAN DRIVE Tasks moved out of the web app Tasks started from the command line, per month, 2026 May: 1 Jun: ~20 Jul: ~110 Aug: ~300 Sep: ~700 Oct: ~6.5k 40% → 19% of tasks started in the web app, May → October ~1.7k vs ~1.2k PRs merged in October from CLI-started vs web-started tasks CLI users went from 1 to 77. Eight of them started 100+ tasks each in October. ### Agents started driving Spectre Four agents used Spectre in different ways. Claude Code split one ticket into 10 milestones and handed 9 to Spectre. Each resulting task had its own sandbox for end-to-end validation. [6:26](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=386s) [Four rows show Claude Code delegating work, Cursor requesting validation, ChatGPT Dots answering questions and an in-house agent steering tasks. Dotted lines lead to Spectre beside a partially filled sandbox grid.](https://assets.ai.engineer/sites/aiecode2025/sha256/936515d1ae446e2f55183c1cadc16766131a75d5ccab146e94da148c9510e7e4.webp) STORY 2 · ANY AGENT CAN DRIVE Agents started driving Spectre Claude Code, in a cloud session Split one ticket into 10, handed 9 to Spectre Cursor Writes the change, asks Spectre to validate it ChatGPT Dots Triages Spectre’s questions, replies through the CLI In-house orchestration agent Starts tasks, steers them, hands work on Spectre Sandboxes “I guess now I just ask Cursor / Claude Code / Dot ... to spam Spectre tasks.” — me, in a DM It split the ticket into ten different milestones. [6:36](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=396s) [Four rows connect Claude Code, Cursor, ChatGPT Dots and an in-house orchestration agent to Spectre. The rows describe delegating work, requesting validation, answering questions and steering tasks. A filled grid represents sandboxes.](https://assets.ai.engineer/sites/aiecode2025/sha256/3ab5e9f9c2597297f17e25828a1d275fff002ae165481701e1948c188b9e234c.webp) STORY 2 · ANY AGENT CAN DRIVE Agents started driving Spectre Claude Code, in a cloud session Split one ticket into 10, handed 9 to Spectre Cursor Writes the change, asks Spectre to validate it ChatGPT Dots Triages Spectre’s questions, replies through the CLI In-house orchestration agent Starts tasks, steers them, hands work on Spectre Sandboxes “I guess now I just ask Cursor / Claude Code / Dot ... to spam Spectre tasks.” — me, in a DM Each has its own sandbox to validate end to end. ### The slowest step wasn’t the model. It was git. Before the agent read a line, Git checkout took about 76 seconds on average. Keeping a bare clone of every repo on each node cut that to about 5 seconds. Each task reuses the local clone and fetches only its branch. [7:21](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=441s) [Submodule initialization takes about 76 seconds and cloning by reference takes about 5 seconds. Each node holds a bare clone, a background job publishes immutable generations, and each task fetches only its branch.](https://assets.ai.engineer/sites/aiecode2025/sha256/e0a9da1571d6f33ff5237d47fc8296f97fb30a03d1ef26ccfaa94fea4a782364.webp) STORY 2 · ANY AGENT CAN DRIVE The slowest step wasn’t the model. It was git. Before · submodule init ~76 s Ready, before the agent read a line After · clone by reference ~5 s Ready THE FIX, IN PLAIN GIT 1 A bare clone of each repo on every node 2 A background job publishes immutable generations 3 Tasks clone by reference and fetch only their branch Every sandbox, every task Worst case before: 700+ s own took over a minute on average. ### One alert, nine hundred agents One bug triggered an alert per workspace. A workflow tagged the agent on each alert, starting about 900 tasks from one channel in roughly four hours. At peak, all 27 nodes were busy. Repeat alerts now join the open investigation. [7:39](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=459s) [A bad release triggered hundreds of workspace alerts. Tagging an agent on each alert generated about 900 tasks from one channel. At peak, all 27 nodes were occupied, and more than 1,000 pod scheduling attempts failed.](https://assets.ai.engineer/sites/aiecode2025/sha256/1890d65ad04f0add95e0c8e57ec9e8449224d330b5dfbe62ceb70394e4b8f515.webp) STORY 3 · THEN IT BROKE One alert, nine hundred agents ~2k tasks started on August 14 ~1.3k of them started by bots 0 of 27 nodes free at the peak 1,000+ failed attempts to schedule a pod A bad release Hundreds of per-workspace alerts A workflow tags the agent on each ~900 tasks from one channel The node pool hits its limit The fix: a repeat alert joins the investigation that’s already open. ### One person can fill the cluster On October 23, one person submitted about 950 tasks in roughly three hours. With no per-person cap on concurrent sandboxes, the cluster hit its node limit. About 380 sandboxes were still pending, and everyone else’s work waited behind them. [8:26](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=506s) [One person created about 950 tasks in three hours. A filled grid shows the node pool at its autoscaling limit. About 380 sandboxes were pending, and other users’ tasks waited behind them.](https://assets.ai.engineer/sites/aiecode2025/sha256/75fe2b957355eb6c5bffbf6d5338af134dd5860b4ec3942e51922d575dd319d6.webp) STORY 3 · THEN IT BROKE One person can fill the cluster ~950 tasks from one person’s command line ~3 hours to create them, October 23 ~380 sandboxes pending at the node limit No cap on concurrent sandboxes per person NODE POOL, AT ITS AUTOSCALING LIMIT PENDING, UNSCHEDULED Created ~950 Pending ~380 Everyone else’s tasks wait behind them That's not a user problem. That's our problem. ### Four days later: a fair launch queue The queue reserves 40 slots for interactive work. Each person can run 25 tasks at once, or up to 50 when there’s room. In a synthetic 300-task burst, the queue cut concurrency from 240 to 20 and p95 task list latency from 1.7 s to 71 ms. [8:58](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=538s) [The queue reserves 40 interactive slots and allows 25 concurrent tasks per person, rising to 50 when capacity permits. In a synthetic 300-task burst, enabling the queue reduced concurrent tasks from 240 to 20 and p95 worker heartbeat latency from 5.3 seconds to 19 milliseconds.](https://assets.ai.engineer/sites/aiecode2025/sha256/06f155a0ab97a0cc1a4391f9d22a9fb860448a48d98f4dc76f6dfc2f78601994.webp) STORY 3 · THEN IT BROKE Four days later: a fair launch queue Interactive work goes first 40 slots are held for people at a keyboard 25 tasks at once per person Borrow up to 50 when the cluster has room Queued tasks show their place In the web app, the CLI and Slack 300-TASK BURST, SYNTHETIC LOAD TEST Queue off Queue on Tasks running at once 240 20 Task list, p95 1.7 s 71 ms Task page, p95 1.7 s 43 ms Worker heartbeat, p95 5.3 s 19 ms No one person, agent or alert gets the whole cluster. ### The next bursts were bigger, and finished On Nov 4, about 4,400 tasks started, more than twice the Aug 14 peak of about 2,000. The share that finished rose from 89% to 94%. The team made several changes, so Wang couldn’t tell which mattered most. [9:33](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=573s) [Bars show tasks started on each month’s busiest day and the share that finished. On August 14, about 2,000 started and 89% finished. On November 4, about 4,400 started and 94% finished. A side panel lists four platform changes.](https://assets.ai.engineer/sites/aiecode2025/sha256/0f5f8af9bd1de6b33c5eecd2e0ad80344df94fcd88a503a543e84f1c4e0c8483.webp) STORY 3 · THEN IT BROKE The next bursts were bigger, and finished Busiest day each month: tasks started, and the share that finished Jul: ~1.1k, 98% Aug 14: ~2k, 89% Sep: ~1.5k, 98% Oct: ~2.6k, 97% Nov 4: ~4.4k, 94% WHAT CHANGED Sep 15 Alert de-duplication Oct 24 Bulk cancel for admins Oct 27 Fair launch queue Oct 29 More sandbox nodes Monthly tasks grew 4.5× since May. Every month, 94–97% of them finished. ### Agents that start work on their own Agents running on a schedule or in response to events started about 2,700 tasks in May 2026 and about 13,000 in October. Alert triage and pull request reviews accounted for about 8,000 of October’s tasks. [10:49](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=649s) [Monthly tasks started by automations rose from about 2,700 in May to 13,000 in October 2026. Alert triage and pull request reviews were the largest October categories. Guardrails included shutdown after five consecutive failures, cost budgets, expiry dates and duplicate-alert suppression.](https://assets.ai.engineer/sites/aiecode2025/sha256/315ac9e20a7145ab9f333568e8c1c60870185452afcee242942cb1e66456ca39.webp) ONE MORE THING Agents that start work on their own Tasks started by automations each month, 2026 May: ~2.7k Jun: ~4.6k Jul: ~8.5k Aug: ~9.6k Sep: ~11k Oct: ~13k WHAT THEY DID IN OCTOBER, ROUGHLY Alert triage: ~4.2k PR reviews: ~3.8k Reports, digests: ~1.7k Tests and CI: ~1.3k Cleanup: ~0.4k ~290 automations ran, owned by ~110 people Guardrails: switched off after 5 failures in a row, cost budgets and expiry dates, repeat alerts de-duplicated. ### What I’m chasing next Wang imagines an always-on agent in every Slack channel. How would teams keep up with merging if agents kept opening pull requests? He wants each PR to show it works before review. At the time, only one merge in 20 carried that proof. [11:39](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=699s) [Three cards ask whether every channel could have an always-available agent, whether each pull request could demonstrate that it works before review, and how to merge when agents continually open new requests.](https://assets.ai.engineer/sites/aiecode2025/sha256/30f7b7839b70b587da7e7d4696cf0eaaea77eb190921ee37986456d5eeeafd2f.webp) ONE MORE THING What I’m chasing next AN AGENT IN EVERY CHANNEL What if every channel had an agent that never logs off? PROOF ON EVERY PR What if every PR proved it works before anyone looked? MERGING AT VOLUME How do you merge when agents never stop opening PRs? ### What I’d do again Joey Wang sees lasting value in the harness that runs, checks and merges code, whichever agent writes it. He’d make every action callable by an agent and cap how much work any one person, agent or alert can start. [12:14](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=734s) [Three recommendations cover owning what happens after code is written, making every action callable by an agent and limiting the work any person, agent or alert can start. The speaker’s personal view is that the lasting work lies in the harness that runs, checks and merges code.](https://assets.ai.engineer/sites/aiecode2025/sha256/4d91438a90202dcb36bbda0fa53c81ead5c743df310bfee2738af5d0004808b8.webp) TAKEAWAYS What I’d do again 1 Own what happens after the code is written 2 Make every action something an agent can call 3 Cap what any one person, agent or alert can start MY PERSONAL VIEW, NOT HARVEY’S POSITION Engineers will write code with whichever agent they like. The lasting work is the harness that runs it, checks it and gets it merged. ### Thank you Joey Wang invites people building similar systems to connect on LinkedIn. He credits Harvey’s own repositories and systems as the source for every number in the talk. [12:47](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=767s) [A dark thank-you slide shows the presenter’s role and a professional profile link. A footer says the talk’s figures came from the company’s own repositories and systems.](https://assets.ai.engineer/sites/aiecode2025/sha256/8708228710a4516bb4e66d18e22c60d652eaea25882fb7437b8ea080e3e40fa9.webp) Thank you Joey Wang · Engineering Lead, Harvey linkedin.com/in/zongziwang Every number in this talk was counted from our own repositories and systems. Harvey ## Full transcript [0:00](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=0s) Let me start with one number from last [0:02](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=2s) Let me start with one number from last [0:02](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=2s) Let me start with one number from last month. On average, more than one agent [0:05](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=5s) month. On average, more than one agent [0:05](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=5s) month. On average, more than one agent has started on our platform every single [0:08](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=8s) has started on our platform every single [0:08](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=8s) has started on our platform every single minute, day and night. [0:11](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=11s) minute, day and night. [0:11](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=11s) minute, day and night. And honestly, the agents were the easy [0:13](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=13s) And honestly, the agents were the easy [0:13](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=13s) And honestly, the agents were the easy part. [0:15](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=15s) part. [0:15](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=15s) part. Hi, I'm Joey. I bu the cloud agent [0:19](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=19s) Hi, I'm Joey. I bu the cloud agent [0:19](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=19s) Hi, I'm Joey. I bu the cloud agent platform we built inside Harvey. [0:21](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=21s) platform we built inside Harvey. [0:21](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=21s) platform we built inside Harvey. After 10,000 PRs, the biggest lesson I [0:25](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=25s) After 10,000 PRs, the biggest lesson I [0:25](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=25s) After 10,000 PRs, the biggest lesson I can share is this. The thing that broke [0:27](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=27s) can share is this. The thing that broke [0:28](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=28s) can share is this. The thing that broke was an agent. It was everything around [0:31](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=31s) was an agent. It was everything around [0:31](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=31s) was an agent. It was everything around them. [0:33](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=33s) them. [0:33](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=33s) them. So, what does one submit add up to? [0:36](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=36s) So, what does one submit add up to? [0:36](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=36s) So, what does one submit add up to? 10,000 PRs that an agent wrote. And it [0:40](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=40s) 10,000 PRs that an agent wrote. And it [0:40](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=40s) 10,000 PRs that an agent wrote. And it isn't just Tavac team playing with a [0:43](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=43s) isn't just Tavac team playing with a [0:43](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=43s) isn't just Tavac team playing with a toy. Last month, about 460 different [0:47](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=47s) toy. Last month, about 460 different [0:47](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=47s) toy. Last month, about 460 different people at Harvey use it. And over the [0:51](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=51s) people at Harvey use it. And over the [0:51](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=51s) people at Harvey use it. And over the summer, it stopped being only about [0:53](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=53s) summer, it stopped being only about [0:53](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=53s) summer, it stopped being only about code. People started asking a data [0:56](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=56s) code. People started asking a data [0:56](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=56s) code. People started asking a data question in Slack and getting answers [0:58](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=58s) question in Slack and getting answers [0:58](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=58s) question in Slack and getting answers back. My favorite recently someone on [1:02](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=62s) back. My favorite recently someone on [1:02](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=62s) back. My favorite recently someone on our marketing team dropped a prototype [1:04](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=64s) our marketing team dropped a prototype [1:04](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=64s) our marketing team dropped a prototype and a talk track into Slack and ask for [1:07](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=67s) and a talk track into Slack and ask for [1:07](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=67s) and a talk track into Slack and ask for a story book for a product video. [1:10](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=70s) a story book for a product video. [1:10](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=70s) a story book for a product video. Once the harness exists, people bring [1:12](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=72s) Once the harness exists, people bring [1:12](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=72s) Once the harness exists, people bring the work that you never planned for. [1:16](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=76s) the work that you never planned for. [1:16](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=76s) the work that you never planned for. So a promise up front. This is not a [1:19](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=79s) So a promise up front. This is not a [1:19](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=79s) So a promise up front. This is not a talk about models, evos or leaderboard. [1:22](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=82s) talk about models, evos or leaderboard. [1:22](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=82s) talk about models, evos or leaderboard. This is talk about hardness, [1:25](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=85s) This is talk about hardness, [1:25](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=85s) This is talk about hardness, odd things around the model, [1:29](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=89s) odd things around the model, [1:29](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=89s) odd things around the model, checking out the ripple, holding the [1:30](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=90s) checking out the ripple, holding the [1:30](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=90s) checking out the ripple, holding the key, opening the PR, getting the merged. [1:34](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=94s) key, opening the PR, getting the merged. [1:34](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=94s) key, opening the PR, getting the merged. When we first wrote about Spectre back [1:36](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=96s) When we first wrote about Spectre back [1:36](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=96s) When we first wrote about Spectre back in April, we put it like this. At some [1:40](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=100s) in April, we put it like this. At some [1:40](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=100s) in April, we put it like this. At some point, a harness become the product. [1:43](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=103s) point, a harness become the product. [1:43](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=103s) point, a harness become the product. Now, whenever I tell people this, I get [1:46](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=106s) Now, whenever I tell people this, I get [1:46](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=106s) Now, whenever I tell people this, I get the same question. Why on earth would [1:48](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=108s) the same question. Why on earth would [1:48](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=108s) the same question. Why on earth would you build this when clock or cursor [1:51](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=111s) you build this when clock or cursor [1:51](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=111s) you build this when clock or cursor already exist? [1:53](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=113s) already exist? [1:53](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=113s) already exist? My answer, we don't compete with them. [1:56](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=116s) My answer, we don't compete with them. [1:56](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=116s) My answer, we don't compete with them. We run them. And here's how I see it as [1:59](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=119s) We run them. And here's how I see it as [1:59](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=119s) We run them. And here's how I see it as my own view. Agents on your laptop are [2:02](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=122s) my own view. Agents on your laptop are [2:02](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=122s) my own view. Agents on your laptop are powerful, but they don't scale. You can [2:05](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=125s) powerful, but they don't scale. You can [2:05](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=125s) powerful, but they don't scale. You can realistically run five to 10 of them. [2:08](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=128s) realistically run five to 10 of them. [2:08](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=128s) realistically run five to 10 of them. Pass that you need a cloud. And cloud [2:11](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=131s) Pass that you need a cloud. And cloud [2:11](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=131s) Pass that you need a cloud. And cloud needs our secrets, our real stack, our [2:14](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=134s) needs our secrets, our real stack, our [2:14](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=134s) needs our secrets, our real stack, our alerts and tickets, one set of rules and [2:18](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=138s) alerts and tickets, one set of rules and [2:18](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=138s) alerts and tickets, one set of rules and enough capacity that nobody gets [2:20](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=140s) enough capacity that nobody gets [2:20](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=140s) enough capacity that nobody gets starved. Those single two owns that for [2:23](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=143s) starved. Those single two owns that for [2:23](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=143s) starved. Those single two owns that for us. So we build layer and that layer [2:26](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=146s) us. So we build layer and that layer [2:26](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=146s) us. So we build layer and that layer exactly what broke. So I want to tell [2:29](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=149s) exactly what broke. So I want to tell [2:29](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=149s) exactly what broke. So I want to tell three stories today. How we stopped [2:32](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=152s) three stories today. How we stopped [2:32](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=152s) three stories today. How we stopped competing on writing code. how we let [2:34](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=154s) competing on writing code. how we let [2:34](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=154s) competing on writing code. how we let any agent drive and were broke once they [2:38](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=158s) any agent drive and were broke once they [2:38](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=158s) any agent drive and were broke once they did. The first story start with the [2:41](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=161s) did. The first story start with the [2:41](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=161s) did. The first story start with the question we keep hearing. [2:44](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=164s) question we keep hearing. [2:44](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=164s) question we keep hearing. More of our engineers were using cursor [2:47](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=167s) More of our engineers were using cursor [2:47](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=167s) More of our engineers were using cursor and they ask how do I get our secret [2:50](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=170s) and they ask how do I get our secret [2:50](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=170s) and they ask how do I get our secret there? How do I test against our real [2:52](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=172s) there? How do I test against our real [2:52](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=172s) there? How do I test against our real stack? We could have rebuilt all that [2:55](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=175s) stack? We could have rebuilt all that [2:56](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=176s) stack? We could have rebuilt all that inside cursor and then again for next to [2:59](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=179s) inside cursor and then again for next to [2:59](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=179s) inside cursor and then again for next to and the one after that. [3:01](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=181s) and the one after that. [3:01](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=181s) and the one after that. Instead, the answer became one sentence. [3:04](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=184s) Instead, the answer became one sentence. [3:04](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=184s) Instead, the answer became one sentence. Just ask cursor to use vector. [3:08](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=188s) Just ask cursor to use vector. [3:08](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=188s) Just ask cursor to use vector. And that's really the whole idea. People [3:10](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=190s) And that's really the whole idea. People [3:10](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=190s) And that's really the whole idea. People will always have a favorite agent and [3:13](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=193s) will always have a favorite agent and [3:13](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=193s) will always have a favorite agent and the favorite keeps changing. So pick [3:15](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=195s) the favorite keeps changing. So pick [3:15](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=195s) the favorite keeps changing. So pick yours. Cloud code, codeex, cursor, trag [3:19](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=199s) yours. Cloud code, codeex, cursor, trag [3:19](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=199s) yours. Cloud code, codeex, cursor, trag your own hands. [3:21](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=201s) your own hands. [3:21](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=201s) your own hands. Start on your laptop, push the session [3:23](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=203s) Start on your laptop, push the session [3:23](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=203s) Start on your laptop, push the session to the cloud, put it back. That's how [3:27](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=207s) to the cloud, put it back. That's how [3:27](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=207s) to the cloud, put it back. That's how Spectre sits underneath all of that. [3:30](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=210s) Spectre sits underneath all of that. [3:30](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=210s) Spectre sits underneath all of that. Because once the code was written, the [3:32](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=212s) Because once the code was written, the [3:32](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=212s) Because once the code was written, the real group starts [3:34](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=214s) real group starts [3:34](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=214s) real group starts picture appears day. Alert files in [3:37](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=217s) picture appears day. Alert files in [3:37](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=217s) picture appears day. Alert files in Slack at 2 in the morning. An agent [3:40](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=220s) Slack at 2 in the morning. An agent [3:40](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=220s) Slack at 2 in the morning. An agent picks up nobody typing. It writes a fix, [3:44](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=224s) picks up nobody typing. It writes a fix, [3:44](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=224s) picks up nobody typing. It writes a fix, spins up real copy of our stack and post [3:48](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=228s) spins up real copy of our stack and post [3:48](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=228s) spins up real copy of our stack and post the proof on the PR. When CI fails, the [3:51](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=231s) the proof on the PR. When CI fails, the [3:52](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=232s) the proof on the PR. When CI fails, the follow-up loop reads the failure and [3:53](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=233s) follow-up loop reads the failure and [3:54](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=234s) follow-up loop reads the failure and pushes another fix while everyone still [3:56](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=236s) pushes another fix while everyone still [3:56](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=236s) pushes another fix while everyone still sleep. By morning, the PR sits in one [4:00](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=240s) sleep. By morning, the PR sits in one [4:00](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=240s) sleep. By morning, the PR sits in one review queue under the same rule as [4:02](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=242s) review queue under the same rule as [4:02](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=242s) review queue under the same rule as other PR whichever agents wrote it. The [4:07](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=247s) other PR whichever agents wrote it. The [4:07](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=247s) other PR whichever agents wrote it. The owner gets a notch and only once your [4:10](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=250s) owner gets a notch and only once your [4:10](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=250s) owner gets a notch and only once your working day starts and once approved a [4:13](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=253s) working day starts and once approved a [4:13](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=253s) working day starts and once approved a agent can automerge it. [4:17](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=257s) agent can automerge it. [4:17](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=257s) agent can automerge it. If you only build one of those stages [4:20](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=260s) If you only build one of those stages [4:20](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=260s) If you only build one of those stages that I described, build a follow-up [4:22](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=262s) that I described, build a follow-up [4:22](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=262s) that I described, build a follow-up loop. We switched on for everyone in [4:25](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=265s) loop. We switched on for everyone in [4:25](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=265s) loop. We switched on for everyone in early October and that months. Four out [4:28](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=268s) early October and that months. Four out [4:28](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=268s) early October and that months. Four out of five agent PRs merged into our [4:30](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=270s) of five agent PRs merged into our [4:30](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=270s) of five agent PRs merged into our backend PR had been through it. [4:35](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=275s) backend PR had been through it. [4:35](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=275s) backend PR had been through it. That's the difference between agent that [4:36](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=276s) That's the difference between agent that [4:36](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=276s) That's the difference between agent that write code and an agent that finished [4:39](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=279s) write code and an agent that finished [4:39](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=279s) write code and an agent that finished the job. [4:41](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=281s) the job. [4:41](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=281s) the job. And then there's the knowledge. Every [4:43](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=283s) And then there's the knowledge. Every [4:43](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=283s) And then there's the knowledge. Every team has that one person who knows [4:45](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=285s) team has that one person who knows [4:45](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=285s) team has that one person who knows exactly how to test the tricky service. [4:48](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=288s) exactly how to test the tricky service. [4:48](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=288s) exactly how to test the tricky service. Uh which flag you never touch. And that [4:51](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=291s) Uh which flag you never touch. And that [4:51](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=291s) Uh which flag you never touch. And that knowledge leaves in their head. of their [4:53](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=293s) knowledge leaves in their head. of their [4:54](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=294s) knowledge leaves in their head. of their laptop. [4:55](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=295s) laptop. [4:55](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=295s) laptop. So we gave it a hope projects [4:59](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=299s) So we gave it a hope projects [4:59](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=299s) So we gave it a hope projects people write down what matters. Adrian [5:01](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=301s) people write down what matters. Adrian [5:01](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=301s) people write down what matters. Adrian write down what they learn. Team pings [5:04](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=304s) write down what they learn. Team pings [5:04](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=304s) write down what they learn. Team pings the skills you trust and there's a [5:06](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=306s) the skills you trust and there's a [5:06](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=306s) the skills you trust and there's a record of who rent what. [5:09](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=309s) record of who rent what. [5:09](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=309s) record of who rent what. Any agent working in the project with [5:11](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=311s) Any agent working in the project with [5:11](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=311s) Any agent working in the project with the same page. My personal view [5:14](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=314s) the same page. My personal view [5:14](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=314s) the same page. My personal view whichever agent you pick the audit trail [5:16](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=316s) whichever agent you pick the audit trail [5:16](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=316s) whichever agent you pick the audit trail and the checks should be the same. [5:19](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=319s) and the checks should be the same. [5:19](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=319s) and the checks should be the same. The second story is about rule. Back in [5:23](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=323s) The second story is about rule. Back in [5:23](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=323s) The second story is about rule. Back in May, if you want a spectctor to do [5:25](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=325s) May, if you want a spectctor to do [5:25](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=325s) May, if you want a spectctor to do something, you open a browser tab or [5:28](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=328s) something, you open a browser tab or [5:28](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=328s) something, you open a browser tab or tacking Slack. The whole month, exactly [5:32](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=332s) tacking Slack. The whole month, exactly [5:32](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=332s) tacking Slack. The whole month, exactly one task came out from terminal. [5:35](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=335s) one task came out from terminal. [5:35](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=335s) one task came out from terminal. So, we made a rule. If a person can [5:37](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=337s) So, we made a rule. If a person can [5:37](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=337s) So, we made a rule. If a person can click it, an agent can call it. [5:41](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=341s) click it, an agent can call it. [5:41](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=341s) click it, an agent can call it. Start a task. Turn on the follow-up loop [5:44](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=344s) Start a task. Turn on the follow-up loop [5:44](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=344s) Start a task. Turn on the follow-up loop on automerge, publish a page, all from [5:47](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=347s) on automerge, publish a page, all from [5:47](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=347s) on automerge, publish a page, all from CLI, [5:49](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=349s) CLI, [5:49](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=349s) CLI, which means your laptop agent can call [5:51](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=351s) which means your laptop agent can call [5:51](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=351s) which means your laptop agent can call it. A agent inside the sandbox can also [5:54](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=354s) it. A agent inside the sandbox can also [5:54](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=354s) it. A agent inside the sandbox can also call it. Even CI can call it too. And [5:57](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=357s) call it. Even CI can call it too. And [5:57](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=357s) call it. Even CI can call it too. And then people ran with it. By October, the [6:01](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=361s) then people ran with it. By October, the [6:01](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=361s) then people ran with it. By October, the command line was starting about 6,500 ts [6:04](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=364s) command line was starting about 6,500 ts [6:04](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=364s) command line was starting about 6,500 ts a month. The web app didn't shrink. [6:08](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=368s) a month. The web app didn't shrink. [6:08](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=368s) a month. The web app didn't shrink. Everyone else just grew around it. [6:11](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=371s) Everyone else just grew around it. [6:11](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=371s) Everyone else just grew around it. Task started from the terminal were [6:13](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=373s) Task started from the terminal were [6:13](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=373s) Task started from the terminal were merging mops and task started in browser [6:16](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=376s) merging mops and task started in browser [6:16](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=376s) merging mops and task started in browser and the new kind of users showed up. [6:19](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=379s) and the new kind of users showed up. [6:19](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=379s) and the new kind of users showed up. People who start hundreds of tasks a [6:21](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=381s) People who start hundreds of tasks a [6:21](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=381s) People who start hundreds of tasks a month and never open the app at all. [6:25](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=385s) month and never open the app at all. [6:25](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=385s) month and never open the app at all. And this is where it gets fun. In [6:27](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=387s) And this is where it gets fun. In [6:27](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=387s) And this is where it gets fun. In October, I had a big task. So I bought [6:31](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=391s) October, I had a big task. So I bought [6:31](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=391s) October, I had a big task. So I bought code session and asked to be the [6:32](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=392s) code session and asked to be the [6:32](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=392s) code session and asked to be the coordinator. [6:34](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=394s) coordinator. [6:34](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=394s) coordinator. It split the tickets into 10 different [6:37](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=397s) It split the tickets into 10 different [6:37](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=397s) It split the tickets into 10 different milestones. K one for me and the rest to [6:41](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=401s) milestones. K one for me and the rest to [6:41](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=401s) milestones. K one for me and the rest to Spectre with hundreds of tasks. Each has [6:44](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=404s) Spectre with hundreds of tasks. Each has [6:44](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=404s) Spectre with hundreds of tasks. Each has its own sandbox to validate end to end. [6:48](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=408s) its own sandbox to validate end to end. [6:48](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=408s) its own sandbox to validate end to end. Then at the end of the October, our it [6:50](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=410s) Then at the end of the October, our it [6:50](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=410s) Then at the end of the October, our it give everyone trag [6:53](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=413s) give everyone trag [6:53](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=413s) give everyone trag now watch Spectre for me. It list the [6:56](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=416s) now watch Spectre for me. It list the [6:56](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=416s) now watch Spectre for me. It list the question Spectre is waiting on draft my [6:58](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=418s) question Spectre is waiting on draft my [6:58](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=418s) question Spectre is waiting on draft my answers and once I say yes, send them [7:01](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=421s) answers and once I say yes, send them [7:01](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=421s) answers and once I say yes, send them back. [7:03](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=423s) back. [7:03](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=423s) back. I actually DM to my friend like that. I [7:06](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=426s) I actually DM to my friend like that. I [7:06](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=426s) I actually DM to my friend like that. I just use cursor clock code or whatever [7:08](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=428s) just use cursor clock code or whatever [7:08](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=428s) just use cursor clock code or whatever to spam spectre. Now, [7:11](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=431s) to spam spectre. Now, [7:11](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=431s) to spam spectre. Now, well, that's the dream, but it's also [7:13](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=433s) well, that's the dream, but it's also [7:13](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=433s) well, that's the dream, but it's also the problem. [7:15](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=435s) the problem. [7:15](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=435s) the problem. A quick detour. When we look at where [7:18](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=438s) A quick detour. When we look at where [7:18](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=438s) A quick detour. When we look at where our task time actually went, the slowest [7:20](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=440s) our task time actually went, the slowest [7:20](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=440s) our task time actually went, the slowest step wasn't model syncing. It was get [7:24](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=444s) step wasn't model syncing. It was get [7:24](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=444s) step wasn't model syncing. It was get just checking out every ripple we own [7:26](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=446s) just checking out every ripple we own [7:26](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=446s) just checking out every ripple we own took over a minute on average. [7:30](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=450s) took over a minute on average. [7:30](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=450s) took over a minute on average. The fix keep a copy of every ripple on [7:33](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=453s) The fix keep a copy of every ripple on [7:33](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=453s) The fix keep a copy of every ripple on every machine and let each task borrow [7:36](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=456s) every machine and let each task borrow [7:36](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=456s) every machine and let each task borrow it. [7:37](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=457s) it. [7:38](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=458s) it. The third story is the one where it [7:40](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=460s) The third story is the one where it [7:40](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=460s) The third story is the one where it broke in mid August. A release goes out [7:44](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=464s) broke in mid August. A release goes out [7:44](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=464s) broke in mid August. A release goes out with a bug. That bug triggers alert that [7:47](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=467s) with a bug. That bug triggers alert that [7:47](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=467s) with a bug. That bug triggers alert that fires once per workspace. So one bug [7:50](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=470s) fires once per workspace. So one bug [7:50](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=470s) fires once per workspace. So one bug become hundreds of alerts all landing in [7:53](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=473s) become hundreds of alerts all landing in [7:53](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=473s) become hundreds of alerts all landing in one single S channel. [7:55](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=475s) one single S channel. [7:55](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=475s) one single S channel. Someone has set up workflow that tag our [7:57](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=477s) Someone has set up workflow that tag our [7:58](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=478s) Someone has set up workflow that tag our agent on every alert and every tag [8:01](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=481s) agent on every alert and every tag [8:01](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=481s) agent on every alert and every tag started task. On normal day those task [8:04](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=484s) started task. On normal day those task [8:04](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=484s) started task. On normal day those task started about 50 task. That day in about [8:08](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=488s) started about 50 task. That day in about [8:08](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=488s) started about 50 task. That day in about four hours they started around 900. [8:12](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=492s) four hours they started around 900. [8:12](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=492s) four hours they started around 900. 900 agents are chasing the same bug [8:14](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=494s) 900 agents are chasing the same bug [8:14](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=494s) 900 agents are chasing the same bug until that wasn't room left for a single [8:17](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=497s) until that wasn't room left for a single [8:17](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=497s) until that wasn't room left for a single note. [8:18](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=498s) note. [8:18](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=498s) note. The fix was almost boring. A repeat [8:21](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=501s) The fix was almost boring. A repeat [8:21](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=501s) The fix was almost boring. A repeat alert now joins the investigation that's [8:23](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=503s) alert now joins the investigation that's [8:23](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=503s) alert now joins the investigation that's already open. [8:25](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=505s) already open. [8:25](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=505s) already open. Two months later, it happened again. And [8:28](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=508s) Two months later, it happened again. And [8:28](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=508s) Two months later, it happened again. And this time, it wasn't a bot. On October, [8:32](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=512s) this time, it wasn't a bot. On October, [8:32](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=512s) this time, it wasn't a bot. On October, the engineer did exactly what we've been [8:35](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=515s) the engineer did exactly what we've been [8:35](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=515s) the engineer did exactly what we've been asking everyone to do. Hand work to [8:38](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=518s) asking everyone to do. Hand work to [8:38](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=518s) asking everyone to do. Hand work to Spectre. [8:40](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=520s) Spectre. [8:40](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=520s) Spectre. About 950 tasks from the command line. [8:43](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=523s) About 950 tasks from the command line. [8:43](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=523s) About 950 tasks from the command line. In about 3 hours, nothing in our [8:46](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=526s) In about 3 hours, nothing in our [8:46](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=526s) In about 3 hours, nothing in our platform can work. The cluster filled up [8:50](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=530s) platform can work. The cluster filled up [8:50](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=530s) platform can work. The cluster filled up and everyone else work sat in line [8:53](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=533s) and everyone else work sat in line [8:53](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=533s) and everyone else work sat in line behind one laptop. That's not a user [8:55](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=535s) behind one laptop. That's not a user [8:55](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=535s) behind one laptop. That's not a user problem. That's our problem. [8:59](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=539s) problem. That's our problem. [8:59](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=539s) problem. That's our problem. 4 days later, we had a real fix. A fair [9:03](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=543s) 4 days later, we had a real fix. A fair [9:03](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=543s) 4 days later, we had a real fix. A fair launch queue. Now starting a task gives [9:06](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=546s) launch queue. Now starting a task gives [9:06](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=546s) launch queue. Now starting a task gives you a place in line and the dispatcher [9:09](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=549s) you a place in line and the dispatcher [9:09](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=549s) you a place in line and the dispatcher started when there's room. People [9:11](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=551s) started when there's room. People [9:12](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=552s) started when there's room. People sitting at keyboard goes first. Everyone [9:15](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=555s) sitting at keyboard goes first. Everyone [9:15](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=555s) sitting at keyboard goes first. Everyone gets a fair share and can borrow more [9:17](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=557s) gets a fair share and can borrow more [9:17](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=557s) gets a fair share and can borrow more than the cluster is quite. [9:19](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=559s) than the cluster is quite. [9:20](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=560s) than the cluster is quite. And whenever you start a task, you can [9:22](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=562s) And whenever you start a task, you can [9:22](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=562s) And whenever you start a task, you can see a placing line. When we hit it with [9:24](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=564s) see a placing line. When we hit it with [9:24](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=564s) see a placing line. When we hit it with a 300 task burst in low test, the task [9:28](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=568s) a 300 task burst in low test, the task [9:28](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=568s) a 300 task burst in low test, the task lit went from almost 2 seconds to 70 [9:30](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=570s) lit went from almost 2 seconds to 70 [9:30](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=570s) lit went from almost 2 seconds to 70 milliseconds. [9:32](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=572s) milliseconds. [9:32](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=572s) milliseconds. So did it work? [9:35](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=575s) So did it work? [9:35](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=575s) So did it work? When it's the busiest day of every week [9:38](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=578s) When it's the busiest day of every week [9:38](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=578s) When it's the busiest day of every week back in August, two certain task and [9:41](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=581s) back in August, two certain task and [9:41](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=581s) back in August, two certain task and about one in 10 didn't finish. [9:45](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=585s) about one in 10 didn't finish. [9:45](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=585s) about one in 10 didn't finish. Then in November 4th, more than twice [9:47](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=587s) Then in November 4th, more than twice [9:47](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=587s) Then in November 4th, more than twice that and much better in the finish line, [9:51](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=591s) that and much better in the finish line, [9:51](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=591s) that and much better in the finish line, but we still have some room to grow. [9:55](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=595s) but we still have some room to grow. [9:55](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=595s) but we still have some room to grow. I will be modest, we change several [9:56](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=596s) I will be modest, we change several [9:56](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=596s) I will be modest, we change several things at once, so I cannot tell you [9:59](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=599s) things at once, so I cannot tell you [9:59](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=599s) things at once, so I cannot tell you which fix mattered the most, but the [10:01](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=601s) which fix mattered the most, but the [10:01](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=601s) which fix mattered the most, but the birds kept getting bigger and the [10:03](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=603s) birds kept getting bigger and the [10:03](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=603s) birds kept getting bigger and the platform kept up. [10:06](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=606s) platform kept up. [10:06](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=606s) platform kept up. And just when that was on the crunch [10:08](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=608s) And just when that was on the crunch [10:08](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=608s) And just when that was on the crunch show, the model neck moves. [10:10](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=610s) show, the model neck moves. [10:10](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=610s) show, the model neck moves. At the end of the October, our dev [10:13](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=613s) At the end of the October, our dev [10:13](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=613s) At the end of the October, our dev platform team raised an alarm. PRS [10:16](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=616s) platform team raised an alarm. PRS [10:16](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=616s) platform team raised an alarm. PRS across the company had jumped around 60% [10:20](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=620s) across the company had jumped around 60% [10:20](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=620s) across the company had jumped around 60% in a single month. A merge that used to [10:23](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=623s) in a single month. A merge that used to [10:23](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=623s) in a single month. A merge that used to take 15 to 30 minutes was taking over an [10:26](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=626s) take 15 to 30 minutes was taking over an [10:26](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=626s) take 15 to 30 minutes was taking over an hour now. And merge Q kept having to be [10:29](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=629s) hour now. And merge Q kept having to be [10:29](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=629s) hour now. And merge Q kept having to be switched off. So now a PR rebase onto [10:32](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=632s) switched off. So now a PR rebase onto [10:32](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=632s) switched off. So now a PR rebase onto main rerun CI and merge if it's green. [10:37](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=637s) main rerun CI and merge if it's green. [10:37](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=637s) main rerun CI and merge if it's green. If main breaks everyone CI breaks with [10:39](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=639s) If main breaks everyone CI breaks with [10:39](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=639s) If main breaks everyone CI breaks with it. Alterately [10:42](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=642s) it. Alterately [10:42](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=642s) it. Alterately we haven't cracked this one yet. [10:45](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=645s) we haven't cracked this one yet. [10:45](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=645s) we haven't cracked this one yet. But before I go there's one more thing. [10:49](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=649s) But before I go there's one more thing. [10:49](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=649s) But before I go there's one more thing. While we were busy fixing all it [10:52](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=652s) While we were busy fixing all it [10:52](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=652s) While we were busy fixing all it something else was quite growing. Most [10:55](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=655s) something else was quite growing. Most [10:55](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=655s) something else was quite growing. Most of what I've been showing you start with [10:57](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=657s) of what I've been showing you start with [10:57](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=657s) of what I've been showing you start with a person asking. [10:59](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=659s) a person asking. [10:59](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=659s) a person asking. The fastest growing source of the work [11:02](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=662s) The fastest growing source of the work [11:02](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=662s) The fastest growing source of the work on Spectre doesn't these are automations [11:06](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=666s) on Spectre doesn't these are automations [11:06](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=666s) on Spectre doesn't these are automations agent that wake up on schedule on an [11:08](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=668s) agent that wake up on schedule on an [11:08](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=668s) agent that wake up on schedule on an event and just do the work tri alert [11:13](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=673s) event and just do the work tri alert [11:13](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=673s) event and just do the work tri alert read the PR write the post. [11:16](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=676s) read the PR write the post. [11:16](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=676s) read the PR write the post. Last month about 110 people had [11:19](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=679s) Last month about 110 people had [11:19](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=679s) Last month about 110 people had automation running and we learned to [11:21](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=681s) automation running and we learned to [11:21](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=681s) automation running and we learned to give them a guardrail on day one. [11:24](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=684s) give them a guardrail on day one. [11:24](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=684s) give them a guardrail on day one. If one fails five times in a row, it [11:26](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=686s) If one fails five times in a row, it [11:26](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=686s) If one fails five times in a row, it switch itself off and it can carry a [11:30](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=690s) switch itself off and it can carry a [11:30](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=690s) switch itself off and it can carry a budget or expiry date because when no [11:33](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=693s) budget or expiry date because when no [11:33](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=693s) budget or expiry date because when no one's watching, an agent can start [11:35](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=695s) one's watching, an agent can start [11:35](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=695s) one's watching, an agent can start thousands of tasks, too. [11:38](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=698s) thousands of tasks, too. [11:38](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=698s) thousands of tasks, too. So, here's what I'm chasing next, and I [11:41](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=701s) So, here's what I'm chasing next, and I [11:41](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=701s) So, here's what I'm chasing next, and I don't have an answer yet, honestly. [11:44](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=704s) don't have an answer yet, honestly. [11:44](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=704s) don't have an answer yet, honestly. What if every Slack channel had its own [11:46](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=706s) What if every Slack channel had its own [11:46](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=706s) What if every Slack channel had its own agent that never locks off? One that [11:49](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=709s) agent that never locks off? One that [11:49](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=709s) agent that never locks off? One that follows competition and start work on [11:51](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=711s) follows competition and start work on [11:51](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=711s) follows competition and start work on its own. What if every pure prove it [11:54](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=714s) its own. What if every pure prove it [11:54](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=714s) its own. What if every pure prove it works before a human even look at it? [11:58](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=718s) works before a human even look at it? [11:58](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=718s) works before a human even look at it? Today only one merge in 20 carries that [12:01](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=721s) Today only one merge in 20 carries that [12:01](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=721s) Today only one merge in 20 carries that kind of proof. [12:03](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=723s) kind of proof. [12:03](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=723s) kind of proof. And how do you merge when agents never [12:06](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=726s) And how do you merge when agents never [12:06](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=726s) And how do you merge when agents never stop opening PRs? [12:08](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=728s) stop opening PRs? [12:08](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=728s) stop opening PRs? Ask me again 6 months. I will tell you [12:11](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=731s) Ask me again 6 months. I will tell you [12:11](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=731s) Ask me again 6 months. I will tell you what broke. If you're building something [12:14](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=734s) what broke. If you're building something [12:14](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=734s) what broke. If you're building something like this when you work, here's three [12:16](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=736s) like this when you work, here's three [12:16](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=736s) like this when you work, here's three things I do again. Oh, what happens [12:19](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=739s) things I do again. Oh, what happens [12:19](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=739s) things I do again. Oh, what happens after code is written. Make Eric action [12:23](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=743s) after code is written. Make Eric action [12:23](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=743s) after code is written. Make Eric action something an agent can call and cap what [12:26](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=746s) something an agent can call and cap what [12:26](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=746s) something an agent can call and cap what an agent, human or alert can start [12:29](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=749s) an agent, human or alert can start [12:29](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=749s) an agent, human or alert can start before they find out what they can do [12:31](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=751s) before they find out what they can do [12:31](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=751s) before they find out what they can do with it. [12:33](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=753s) with it. [12:33](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=753s) with it. And one last thought, that's my own [12:35](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=755s) And one last thought, that's my own [12:36](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=756s) And one last thought, that's my own view. Engineers can write code with [12:39](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=759s) view. Engineers can write code with [12:39](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=759s) view. Engineers can write code with whichever agent they like. The work [12:42](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=762s) whichever agent they like. The work [12:42](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=762s) whichever agent they like. The work that's lost is the harness that runs it, [12:44](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=764s) that's lost is the harness that runs it, [12:44](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=764s) that's lost is the harness that runs it, checks it, and get it merged. [12:48](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=768s) checks it, and get it merged. [12:48](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=768s) checks it, and get it merged. Thank you so much for watching. [12:51](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=771s) Thank you so much for watching. [12:51](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=771s) Thank you so much for watching. Come find me on linking if you are [12:53](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=773s) Come find me on linking if you are [12:53](https://www.youtube.com/watch?v=ktbAVYsAxpk&t=773s) Come find me on linking if you are building similar things.