# The harness has to finish the job

Source: https://www.youtube.com/watch?v=ZZ19zhYnDKY

Machine-generated transcript; recognition errors remain. Source-recording timestamps. Boundary-crossing subtitle ends are clipped to this session.

[01:37:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5870s) >> Let me start with one number from last

[01:37:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5872s) month. On average, more than one agent

[01:37:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5875s) has started on our platform every single

[01:37:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5878s) minute, day and night.

[01:38:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5881s) And honestly, the agents were the easy

[01:38:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5883s) part.

[01:38:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5885s) Hi, I'm Joey. I bu the cloud agent

[01:38:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5889s) platform we built inside Harvey.

[01:38:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5891s) After 10,000 PRs, the biggest lesson I

[01:38:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5895s) can share is this. The thing that broke

[01:38:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5897s) was an agent. It was everything around

[01:38:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5900s) them.

[01:38:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5903s) So what does one submit add up to?

[01:38:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5906s) 10,000 PRs that an agent wrote. And it

[01:38:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5910s) isn't just Tax team playing with a toy.

[01:38:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5914s) Last month about 460 different people at

[01:38:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5918s) Harvey use it. And over the summer it

[01:38:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5922s) stopped being only about code. People

[01:38:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5924s) started asking a data question in Slack

[01:38:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5927s) and getting answers back. My favorite

[01:38:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5931s) recently someone on our marketing team

[01:38:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5933s) drop a prototype and a talk track into

[01:38:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5935s) Slack and ask for a story book for a

[01:38:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5938s) product video. Once the harness exists,

[01:39:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5942s) people bring the work that you never

[01:39:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5944s) plan for.

[01:39:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5946s) So a promise up front. This is not a

[01:39:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5949s) talk about models, Evos or leaderboard.

[01:39:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5952s) This is talk about hardness,

[01:39:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5955s) odd things around the model,

[01:39:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5958s) checking out the ripple, holding the

[01:39:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5960s) key, opening the PR, getting the merged.

[01:39:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5964s) When we first wrote about Spectre back

[01:39:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5966s) in April, we put it like this. At some

[01:39:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5970s) point, a harness become the product.

[01:39:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5973s) Now, whenever I tell people this, I get

[01:39:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5976s) the same question. Why on earth would

[01:39:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5978s) you build this when clock or cursor

[01:39:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5981s) already exist?

[01:39:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5983s) My answer, we don't compete with it. We

[01:39:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5986s) run them. And here's how I see it as my

[01:39:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5989s) own view. Agents on your laptop are

[01:39:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5992s) powerful, but they don't scale. You can

[01:39:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5995s) realistically run five to 10 of them.

[01:39:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=5998s) Pass that you need a cloud. And cloud

[01:40:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6001s) needs our secrets, our real stack, our

[01:40:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6004s) alerts and tickets, one set of rules and

[01:40:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6008s) enough capacity that nobody gets

[01:40:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6010s) starved. Those single two owns that for

[01:40:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6013s) us. So we build layer and that layer

[01:40:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6016s) exactly what broke. So I want to tell

[01:40:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6019s) three stories today. How we stop

[01:40:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6022s) competing on writing code. How we let

[01:40:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6024s) any agent drive and were broke once they

[01:40:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6028s) did. The first story start with the

[01:40:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6031s) question we keep hearing.

[01:40:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6034s) More of our engineers were using cursor

[01:40:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6037s) and they ask how do I get our secret

[01:40:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6040s) there? How do I test against our real

[01:40:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6042s) stack? We could have rebuilt all that

[01:40:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6046s) inside cursor and then again for next

[01:40:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6048s) two and the one after that.

[01:40:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6051s) Instead answer became one sentence just

[01:40:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6055s) ask cursor to use spectre

[01:40:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6058s) and that's really the whole idea. People

[01:41:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6060s) will always have a favorite agent and

[01:41:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6063s) the favorite keeps changing. So pick

[01:41:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6065s) yours clock codeex cursor dot slack are

[01:41:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6069s) your own hands.

[01:41:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6071s) Start on your laptop push the session to

[01:41:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6074s) the cloud. Put it back. That's how

[01:41:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6077s) spectra sits underneath all of that.

[01:41:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6080s) Because once the code was written the

[01:41:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6082s) real work starts

[01:41:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6084s) picture appears day alert fires in Slack

[01:41:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6088s) at 2 in the morning. An agent picks up

[01:41:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6091s) nobody typing. It writes a fix, spins up

[01:41:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6095s) real copy of our stack and post the

[01:41:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6098s) proof on the PR.

[01:41:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6100s) When C fails, the follow-up loop reads

[01:41:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6103s) the failure and pushes another fix while

[01:41:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6105s) everyone still sleep. By morning the PR

[01:41:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6109s) sits in one review queue under the same

[01:41:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6112s) rule as other PR whichever agents wrote

[01:41:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6116s) it. The owner gets a notch and only once

[01:41:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6119s) they are working day starts and once

[01:42:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6122s) approved a agent can automerge it.

[01:42:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6127s) If you only build one of those stages

[01:42:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6130s) that I described build a follow-up loop.

[01:42:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6133s) We switched it on for everyone in early

[01:42:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6135s) October and that months. Four out of

[01:42:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6138s) five agent PRs merge into our backend PR

[01:42:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6142s) had been through it.

[01:42:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6144s) That's the difference between agent that

[01:42:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6146s) write code and an agent that finished

[01:42:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6149s) the job.

[01:42:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6151s) And then there's the knowledge. Every

[01:42:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6153s) team has that one person who knows

[01:42:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6155s) exactly how to test the tricky service

[01:42:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6158s) which flag you never touch. And that

[01:42:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6161s) knowledge leaves in their head. of their

[01:42:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6164s) laptop.

[01:42:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6165s) So we gave it a hope projects.

[01:42:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6169s) People write down what matters. Adrian

[01:42:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6171s) write down what they learn. Team pings

[01:42:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6174s) the skills you trust and there's a

[01:42:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6176s) record of who rent what. Any agent

[01:42:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6179s) working in the project read the same

[01:43:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6181s) page. My personal view whichever agent

[01:43:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6184s) you pick the audit trail and the tracks

[01:43:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6187s) should be the same.

[01:43:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6189s) The second story is about rule.

[01:43:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6193s) Back in May, if you want a spectctor to

[01:43:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6195s) do something, you open a browser tab or

[01:43:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6198s) tacking Slack. The whole mouse, exactly

[01:43:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6202s) one task came out from terminal.

[01:43:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6205s) So, we made a rule. If a person can

[01:43:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6207s) click it, an agent can call it

[01:43:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6210s) start a task, turn on the follow-up

[01:43:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6213s) loop, arm, automerge, publish a page,

[01:43:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6216s) all from CLI,

[01:43:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6219s) which means your laptop agent can call

[01:43:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6221s) it A agent inside the sandbox can also

[01:43:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6224s) call it. Even CI can call it too. And

[01:43:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6227s) then people ran with it. By October, the

[01:43:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6231s) command line was starting about 6500 ts

[01:43:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6234s) a month. The web app didn't shrink.

[01:43:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6238s) Everyone else just grew around it. Task

[01:44:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6241s) started from the terminal were merging

[01:44:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6244s) and task started in browser. And the new

[01:44:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6246s) kind of user showed up. people who start

[01:44:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6250s) hundreds of tasks a month and never open

[01:44:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6252s) the app at all.

[01:44:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6255s) And this is where it gets fun. In

[01:44:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6257s) October, I had a big task. So I bought

[01:44:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6261s) code session and asked to be the

[01:44:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6262s) coordinator.

[01:44:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6264s) It split the tickets into 10 different

[01:44:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6267s) milestones. Get one for me and the rest

[01:44:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6270s) to spectre with hundreds of tasks.

[01:44:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6274s) Each has its own sandbox to validate end

[01:44:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6276s) to end. Then at the end of the October,

[01:44:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6280s) our it give everyone tragic dots. Mind

[01:44:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6283s) now watch Spectre for me. It list the

[01:44:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6286s) question Spectre is waiting on. Draft my

[01:44:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6288s) answers and once I say yes, send them

[01:44:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6291s) back.

[01:44:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6293s) I actually DM to my friend like that. I

[01:44:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6296s) just use cursor clock code or whatever

[01:44:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6298s) to spam spectre. Now

[01:45:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6301s) well that's the dream, but it's also the

[01:45:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6303s) problem. A quick detour. When we look at

[01:45:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6307s) where our task time actually went, the

[01:45:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6310s) slowest step wasn't model syncing. It

[01:45:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6313s) was get just checking out every ripple

[01:45:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6315s) we own took over a minute on average.

[01:45:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6320s) The fix keep a copy of every ripple on

[01:45:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6323s) every machine and let each test borrow

[01:45:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6326s) it.

[01:45:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6327s) The third story is the one where it

[01:45:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6330s) broke in mid August. A release goes out

[01:45:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6334s) with a bug. That bug triggers alert that

[01:45:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6337s) fires once per workspace. So one bug

[01:45:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6340s) become hundreds of alerts all landing in

[01:45:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6343s) one single sack channel. Someone has set

[01:45:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6346s) up workflow that tag our agent on every

[01:45:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6348s) alert and every tag started task. On

[01:45:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6353s) normal day those task started about 50

[01:45:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6355s) task. That day in about 4 hours they

[01:45:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6359s) started around 900

[01:46:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6362s) 900 agents all chasing the same bug

[01:46:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6364s) until that wasn't room left for a single

[01:46:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6366s) note.

[01:46:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6368s) The fix was almost boring. A repeat

[01:46:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6371s) alert now joins the investigation that's

[01:46:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6373s) already open.

[01:46:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6375s) Two months later it happened again and

[01:46:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6378s) this time it wasn't a bot. On October,

[01:46:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6382s) the engineer did exactly what we've been

[01:46:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6385s) asking everyone to do. Hand work to

[01:46:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6388s) Spectre.

[01:46:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6390s) About 950 test from the command line in

[01:46:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6394s) about 3 hours. Nothing in our platform

[01:46:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6396s) can work. The cluster filled up and

[01:46:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6400s) everyone else work set in line behind

[01:46:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6403s) one laptop. That's not a user problem.

[01:46:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6406s) That's our problem.

[01:46:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6409s) Four days later, we had a real fix. A

[01:46:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6412s) fail launch queue. Now starting a task

[01:46:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6416s) gives you a place in line. And the

[01:46:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6418s) dispatcher started when there's room.

[01:47:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6421s) People sitting at a keyboard goes first.

[01:47:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6424s) Everyone gets a fair share and can

[01:47:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6426s) borrow more than the cluster is quite.

[01:47:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6429s) And whenever you start a task, you can

[01:47:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6432s) see our placing line. When we hit it

[01:47:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6434s) with a 300 task bursting low test, the

[01:47:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6438s) task lit went from almost two second to

[01:47:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6440s) 70 milliseconds.

[01:47:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6442s) So did it work?

[01:47:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6445s) When it's the busiest day of every week

[01:47:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6448s) back in August, two certain task and

[01:47:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6451s) about one in hand didn't finish.

[01:47:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6455s) Then in November 4th, more than twice

[01:47:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6457s) that and much better in the finish line,

[01:47:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6461s) but we still have some room to grow.

[01:47:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6465s) I will be modest, we change several

[01:47:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6466s) things at once, so I cannot tell you

[01:47:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6469s) which fix mattered the most, but the

[01:47:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6471s) birds kept getting bigger and the

[01:47:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6473s) platform kept up.

[01:47:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6476s) And just when that was on the crunch,

[01:47:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6478s) the model neck moves.

[01:48:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6480s) At the end of October, our dev platform

[01:48:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6484s) team raised an alarm. PRS across the

[01:48:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6487s) company had jumped at around 60% in a

[01:48:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6490s) single month. A merge that used to take

[01:48:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6493s) 15 to 30 minutes was taking over an hour

[01:48:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6496s) now. And merge Q kept having to be

[01:48:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6499s) switched off. So now a PR rebase onto

[01:48:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6502s) main rerun CI and merge if it's green.

[01:48:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6507s) If main breaks, everyone CI breaks with

[01:48:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6509s) it. Alterately,

[01:48:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6512s) we haven't cracked this one yet.

[01:48:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6515s) But before I go, there's one more thing.

[01:48:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6519s) While we were busy fixing all it,

[01:48:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6522s) something else was quite growing. Most

[01:48:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6525s) of what I've been showing you start with

[01:48:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6527s) a person asking

[01:48:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6529s) the fastest growing source of the work

[01:48:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6532s) on Spectre doesn't. These are

[01:48:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6535s) automations agent that wake up on

[01:48:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6537s) schedule on an event and just do the

[01:49:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6540s) work tri alert review the PR write the

[01:49:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6544s) post.

[01:49:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6546s) Last month about 110 people had

[01:49:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6549s) automation running and we learned to

[01:49:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6551s) give them a guardrail on day one. If one

[01:49:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6554s) fails five times in a row, it switch

[01:49:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6557s) itself off and it can carry a budget or

[01:49:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6561s) expiry date because when no one's

[01:49:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6563s) watching, an agent can start thousands

[01:49:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6566s) of tasks, too.

[01:49:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6568s) So, here's what I'm chasing next, and I

[01:49:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6571s) don't have an answer yet, honestly.

[01:49:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6574s) What if every Slack channel had its own

[01:49:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6576s) agent that never locks off? One that

[01:49:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6579s) follows competition and start work on

[01:49:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6581s) its own. What if every pure proof it

[01:49:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6584s) works before a human even look at it?

[01:49:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6588s) Today only one merge in 20 carries that

[01:49:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6591s) kind of proof.

[01:49:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6593s) And how do you merge when agents never

[01:49:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6596s) stop opening PRs?

[01:49:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6598s) Ask me again six months. I will tell you

[01:50:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6601s) what broke. If you're building something

[01:50:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6604s) like this when you work, here's three

[01:50:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6606s) things I do again. Oh, what happens

[01:50:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6609s) after code is written. Make Eric action

[01:50:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6612s) something an agent can call and cap what

[01:50:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6616s) an agent, human or alert can start

[01:50:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6619s) before they find out what they can do

[01:50:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6621s) with it.

[01:50:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6623s) And one last thought, that's my own

[01:50:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6626s) view. Engineers can write code with

[01:50:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6629s) whichever agent they like. The work

[01:50:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6632s) that's lost is the harness that runs it,

[01:50:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6634s) checks it, and get it merged.

[01:50:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6638s) Thank you so much for watching.

[01:50:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6641s) Come find me on linking if you are

[01:50:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=6643s) building similar things.