# From answering the page to sharing the on-call rotation

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.

[01:58:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7118s) and Vlad. We're excited to share our

[01:58:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7120s) journey as well as our learnings around

[01:58:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7123s) building an AI on call teammate at Meta

[01:58:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7126s) that now runs real on call rotations

[01:58:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7128s) every day and around the clock.

[01:58:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7131s) We are part of the engineering team that

[01:58:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7133s) builds the tooling platform and

[01:58:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7135s) infrastructure for incident response on

[01:58:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7138s) call experience and mitigations at meta.

[01:59:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7141s) That space is evolving rapidly with the

[01:59:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7144s) increase in AI capabilities.

[01:59:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7147s) AI changed how fast we write code. About

[01:59:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7150s) 3/4 of new code is now written with AI.

[01:59:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7153s) And more AI generated changes reach

[01:59:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7156s) production every single day. And some of

[01:59:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7159s) it can cause production failures just

[01:59:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7162s) like any other code change can.

[01:59:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7165s) But on call did not speed up with it.

[01:59:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7168s) Highquality code can pass review and

[01:59:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7171s) still break in production. and when it

[01:59:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7173s) does, someone still gets paged at 2 a.m.

[01:59:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7176s) The bottleneck moved from writing code

[01:59:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7178s) to operating it in production. And on

[01:59:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7182s) call is where that operational load

[01:59:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7184s) lands. So we asked if AI is writing the

[01:59:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7188s) code, can AI also help operate it?

[01:59:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7193s) And we think it can. Now let's quickly

[01:59:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7197s) run through one of those on call pages

[01:59:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7199s) with a simplified example.

[02:00:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7202s) It's 2:00 a.m. and 12 alerts are firing.

[02:00:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7205s) Same night, it can go two different

[02:00:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7207s) ways. On the left is how most of us do

[02:00:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7209s) this today. And notice we already have

[02:00:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7212s) good AI. First, you sort the alerts by

[02:00:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7215s) yourself. Checkout, payments, disk,

[02:00:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7218s) which one to pick first. Let's say you

[02:00:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7219s) pick checkout. Then you ask your

[02:00:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7222s) assistant why P99 latency is high.

[02:00:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7226s) It's genuinely useful. It pulls the

[02:00:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7227s) metrics and finds the deployment at 149.

[02:00:31](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7231s) And when you ask, it reads what changed

[02:00:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7234s) and checks the runbook.

[02:00:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7236s) It figured out that the time config was

[02:00:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7238s) changed to a lower value. And based on

[02:00:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7241s) the runbook, it suggests to roll back,

[02:00:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7244s) which is all very helpful. However, if

[02:00:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7247s) you look at the shape of it, it waited

[02:00:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7249s) to be asked.

[02:00:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7251s) You picked every next step. You were the

[02:00:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7254s) glue between the questions and once it

[02:00:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7258s) answered, it went dormant.

[02:01:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7260s) At 239, you know the cause, but the roll

[02:01:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7265s) back, the changes, and the follow-ups

[02:01:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7267s) are still yours.

[02:01:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7270s) Now, on the right, the same 12 alerts,

[02:01:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7274s) but an AI teammate started before you

[02:01:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7277s) woke up. It had already ranked them.

[02:01:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7281s) Checkouts first, payments merged in as

[02:01:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7284s) the same cause, and the other 10 handled

[02:01:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7286s) separately.

[02:01:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7288s) It read the dashboards, difficed

[02:01:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7293s) two past incidents.

[02:01:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7295s) You get a hypothesis, a 30-cond way to

[02:01:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7299s) verify it, and a fix waiting for your

[02:01:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7302s) review. At 2011, you verified it and

[02:01:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7306s) you're back to sleep. That gap is what

[02:01:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7309s) this talk is about.

[02:01:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7313s) So why were you the glue?

[02:01:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7316s) Every team was bolting automation onto

[02:01:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7318s) on call one silo at a time. An alert

[02:02:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7323s) script for alerts, a triage bot for

[02:02:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7325s) tickets, a deploy bot for deploys, a fix

[02:02:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7328s) it agent for incidents.

[02:02:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7330s) Each one solves its slice, keeps its own

[02:02:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7333s) context, and forgets it. Now replaying

[02:02:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7336s) your 2 AM page, the alert script saw P99

[02:02:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7340s) latency above 2 seconds. The triage bot

[02:02:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7343s) saw three tickets saying checkout is

[02:02:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7345s) slow. The deploy bot knew about a config

[02:02:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7348s) change at 149 and the incident history

[02:02:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7350s) had two similar past incidents. Every

[02:02:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7353s) clue existed,

[02:02:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7355s) but nothing connected them. So the on

[02:02:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7359s) call became the glue exactly

[02:02:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7362s) as we saw on the left side of the

[02:02:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7364s) previous slide.

[02:02:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7366s) Our bet was one AI teammate per rotation

[02:02:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7370s) instead.

[02:02:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7372s) One loop across every signal, one team

[02:02:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7375s) memory that learns every shift, its own

[02:02:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7378s) identity so it acts as itself and a

[02:03:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7381s) human gate so you approve instead of

[02:03:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7383s) operate. But giving one agent all that

[02:03:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7387s) context and reach raises the stakes.

[02:03:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7389s) Which is why autonomy has to be earned.

[02:03:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7395s) Under the hood, the teammate or the AI

[02:03:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7398s) teammate is the simplest agent. There is

[02:03:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7401s) model using tools in a loop acting on an

[02:03:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7404s) environment and reading the feedback.

[02:03:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7407s) Everything else in this talk is what we

[02:03:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7409s) [music] build around that loop to make

[02:03:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7412s) it safe to leave it running. And for

[02:03:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7415s) that I will hand over to Swapnikai and

[02:03:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7417s) Vlad next to get into the actual details

[02:03:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7420s) of how it works.

[02:03:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7423s) To achieve this we started with an

[02:03:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7425s) autonomy ladder. Think about how you

[02:03:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7427s) onboard a new engineer to on call. First

[02:03:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7430s) the shadow. The investigators suggest

[02:03:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7433s) but they don't touch production. Then

[02:03:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7436s) the pair the right to fix and another

[02:03:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7439s) engineer approves it. With time they

[02:04:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7441s) gain knowledge. We give our agent

[02:04:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7444s) exactly the same ladder. At level one,

[02:04:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7447s) it investigates and tells you what it

[02:04:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7449s) found. It changes nothing. At level two,

[02:04:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7452s) it acts with approval. It drafts the

[02:04:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7454s) fix, but nothing reaches production

[02:04:16](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7456s) until a human approves. At level three,

[02:04:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7459s) it clears lower criticality work on its

[02:04:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7462s) own and only interrupts you for what

[02:04:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7464s) matters. High criticality work stays

[02:04:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7467s) with humans.

[02:04:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7469s) So what earns the next run? We use a

[02:04:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7472s) simple test. How costly is a mistake and

[02:04:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7475s) how easy is it to catch? So an action

[02:04:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7478s) climbs only when mistakes are cheap.

[02:04:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7481s) They're caught by tests, CI or review.

[02:04:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7485s) and the agent already has a track record

[02:04:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7487s) on that exact action. That's why the

[02:04:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7490s) agent can close a noisy alert on its own

[02:04:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7493s) while a production mitigation still

[02:04:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7495s) waits for a human.

[02:04:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7498s) At level one, the investigation is

[02:05:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7500s) layered and stops at the first confident

[02:05:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7502s) answer. The agent investigated the issue

[02:05:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7505s) using the team's memory and runbook. It

[02:05:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7508s) also shared the root cause analysis, but

[02:05:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7510s) human had to apply the fix.

[02:05:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7513s) At level two, it writes the fix with the

[02:05:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7517s) root cause and a test plan. It answers

[02:05:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7520s) reviewer's comments and fixes CI, but

[02:05:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7523s) nothing reaches production until a human

[02:05:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7525s) approves.

[02:05:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7526s) Most drafts don't land, and that's fine.

[02:05:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7530s) Review is the filter. At both levels,

[02:05:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7533s) the agent does the homework, but you

[02:05:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7536s) still had to grade every single paper.

[02:05:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7539s) The help got much better, but the tasks

[02:05:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7542s) requiring human input didn't get

[02:05:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7544s) smaller. Level three shrinks the pile.

[02:05:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7550s) Look at this example. 23 items came in,

[02:05:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7553s) but only three need you. On the right,

[02:05:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7557s) the agent handled the long tail, noisy

[02:05:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7559s) alerts, duplicates, routine fixes that

[02:06:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7562s) went through normal review and CI. Two

[02:06:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7565s) more are still being worked. You don't

[02:06:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7568s) have to drive each one to completion. In

[02:06:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7571s) the middle is your much smaller decision

[02:06:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7573s) queue. Each item makes clear what it

[02:06:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7577s) needs from you with a run this button

[02:06:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7579s) that has a call to action. And remember

[02:06:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7582s) our 2 a.m. check out page that still

[02:06:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7585s) stays human owned. The agent brings the

[02:06:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7587s) investigation. The human owns the

[02:06:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7590s) critical decision. It's the new high

[02:06:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7592s) ladder again. Trusted to handle routine

[02:06:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7595s) work, ready with the homework when it

[02:06:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7597s) needs help. Not just less work per item,

[02:06:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7600s) but fewer items waiting on you.

[02:06:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7605s) That board is where the work lives. But

[02:06:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7607s) nobody stares at a board all the time,

[02:06:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7610s) right? So the agent is also available

[02:06:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7612s) across many surfaces like chat and

[02:06:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7615s) native incident surfaces. At 9 every

[02:06:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7618s) morning, it posts a digest. 23 came in,

[02:07:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7622s) three need you, 18 handled. Each thing

[02:07:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7626s) that needs you come with a one-click

[02:07:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7628s) next step. So approving the quota bump

[02:07:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7632s) takes one tap from your phone. The 18 it

[02:07:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7635s) handled are summarized, not listed. And

[02:07:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7638s) the two that are still in progress tell

[02:07:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7640s) you what it's doing about them. And when

[02:07:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7643s) you want to dig in, you just reply in

[02:07:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7645s) the thread. As you see on the right

[02:07:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7648s) here, you open the checkout incident and

[02:07:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7650s) ask for the evidence. It already has it.

[02:07:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7654s) Errors start at 8:41, 2 minutes after a

[02:07:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7657s) config push cut the payment client

[02:07:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7659s) timeout. The same pattern as two past

[02:07:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7661s) incidents. You ask a follow-up, is it

[02:07:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7664s) only EU? And it answers with a

[02:07:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7666s) breakdown. It drafts the roll back. You

[02:07:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7669s) approve it right there and errors start

[02:07:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7672s) recovering. You never opened a dashboard

[02:07:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7674s) and you never had to explain the context

[02:07:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7677s) because it was already there. That's

[02:07:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7679s) what a teammate feels like. A morning

[02:08:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7681s) hand off a colleague you can ask.

[02:08:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7686s) Now under the hood of this AI teammate,

[02:08:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7689s) it's AI automations. The first

[02:08:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7692s) investigates on every new signal auto

[02:08:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7695s) triage classifies it, checks its

[02:08:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7698s) criticality and kicks off the right

[02:08:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7700s) agent. An alert goes to the incident

[02:08:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7703s) investigator, a task to the team's own

[02:08:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7705s) agent, a test failure to a coding agent,

[02:08:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7709s) a look up to memory and code search.

[02:08:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7712s) The second keeps the system healthy.

[02:08:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7715s) Unattended automation fails quietly. So

[02:08:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7719s) on a schedule, it looks for stuck or

[02:08:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7721s) failed runs, retrace and repairs broken

[02:08:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7724s) schedules, and it only pages a human if

[02:08:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7727s) it can't fix it itself.

[02:08:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7730s) Nothing stalls silently.

[02:08:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7733s) The third hands off at every shift

[02:08:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7736s) change and every morning at 9, it

[02:08:58](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7738s) summarizes the shift. What needs you

[02:09:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7740s) with the fix ready and what it handled

[02:09:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7743s) on its own posted to the team chat. Not

[02:09:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7746s) only in the morning, but every urgent

[02:09:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7748s) issue that needs user attention gets

[02:09:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7750s) notified in chat with findings and draft

[02:09:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7753s) fixes. It also tags the next on call in

[02:09:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7755s) the chat during handoff. The entire

[02:09:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7758s) cycle still continues

[02:09:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7763s) and because it belongs to the team, it

[02:09:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7766s) runs the agents and the skills the team

[02:09:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7768s) already trusts. That's the slot in the

[02:09:30](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7770s) middle. Bring your own agent, add your

[02:09:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7773s) own skills. If a team already has an

[02:09:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7776s) agent that has the team's runbooks and

[02:09:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7778s) the knowledge base, the automations

[02:09:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7780s) leverage it. Alongside that, it has our

[02:09:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7783s) own incident investigator and a deep

[02:09:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7785s) dive coding session.

[02:09:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7788s) Work comes in from alerts, tickets,

[02:09:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7791s) incidents, team posts, and chat. The

[02:09:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7795s) agent itself is generic. What makes it

[02:09:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7797s) good at an on call rotation is the

[02:09:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7799s) team's agents skills, memory, and

[02:10:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7802s) playbook. And every run is logged, so

[02:10:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7805s) anyone on the team can inspect the

[02:10:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7808s) trajectory or resume the conversation.

[02:10:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7811s) Now, I'm handing off to my teammate Vlad

[02:10:13](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7813s) to talk more on the shared agent.

[02:10:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7819s) An uncle agent can't belong to just a

[02:10:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7821s) single person because the uncle itself

[02:10:23](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7823s) doesn't. So it belongs to the team.

[02:10:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7826s) There's one teammate per rotation on one

[02:10:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7829s) board. Wherever's on call can ask it to

[02:10:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7832s) correct it or pick up its work mid

[02:10:33](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7833s) investigation. It carries the team's

[02:10:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7836s) uncle expertise, the playbook, lord

[02:10:38](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7838s) profiles and past decisions from one

[02:10:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7840s) shift to the next one the way a good

[02:10:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7842s) handoff should. Here's what that looks

[02:10:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7845s) like. Monday at 2 am, Sam is on call and

[02:10:48](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7848s) tells it remember if checkout time is

[02:10:51](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7851s) out after a config push, roll back the

[02:10:53](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7853s) config first. It saves it to the memory.

[02:10:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7857s) Three days later, Jordan is on call hits

[02:11:00](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7860s) the exact same alert and just asks for a

[02:11:03](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7863s) fix. The agent answers with some fix,

[02:11:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7866s) says where it came from, and has the

[02:11:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7868s) roll back drafted. Teach it once, the

[02:11:11](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7871s) next phone call gets it.

[02:11:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7874s) And it access the team its own identity

[02:11:17](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7877s) owned by the rotation with only the

[02:11:19](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7879s) access the team is comfortable granting

[02:11:21](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7881s) it.

[02:11:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7884s) And it learns from every uncle shift.

[02:11:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7886s) Handoffs, decisions, and corrections all

[02:11:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7889s) flow into memory. So ship number 30 is

[02:11:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7892s) better than the ship number one. The

[02:11:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7894s) same way a new uncle gets better every

[02:11:36](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7896s) week. But memory erodess over time. So

[02:11:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7900s) ours has to earn its place. It retrieves

[02:11:43](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7903s) semantically uses what it finds, scores

[02:11:46](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7906s) where it actually helped and promotes or

[02:11:49](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7909s) retires entries automatically.

[02:11:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7912s) When someone says that was wrong, the

[02:11:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7914s) item reopens and the correction becomes

[02:11:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7916s) memory.

[02:11:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7919s) Three things made this work beyond our

[02:12:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7921s) team.

[02:12:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7922s) First, a loop that never stops

[02:12:05](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7925s) listening. Every piece of feedback or

[02:12:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7927s) course correction reopens the item and

[02:12:09](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7929s) ultimately gets absorbed into memory.

[02:12:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7932s) Feedback is an input to the system, not

[02:12:15](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7935s) a story read once a quarter. Second,

[02:12:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7938s) customize instead of forking. Teams

[02:12:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7942s) bring their own playbooks and skill

[02:12:24](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7944s) workflows, and it's still the same

[02:12:26](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7946s) teammate underneath.

[02:12:28](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7948s) Third, what we call one rotation to end.

[02:12:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7952s) Identity, config, and memory are per

[02:12:34](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7954s) team. So, add a rotation becomes a

[02:12:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7957s) setup, not a new project.

[02:12:39](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7959s) The agent is generic while the expertise

[02:12:41](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7961s) is owned by the team.

[02:12:45](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7965s) So what has changed for the people on

[02:12:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7967s) call? Well, first their job shifted from

[02:12:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7970s) driving deep investigations to making

[02:12:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7972s) key decisions. They start from a

[02:12:54](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7974s) hypothesis with linked evidence, not a

[02:12:56](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7976s) blank task or dashboard.

[02:12:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7979s) Second, the load is getting lighter.

[02:13:02](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7982s) Early signs of on call load reduction

[02:13:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7984s) are promising for teams that run it

[02:13:06](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7986s) automatic autonomously and we want to

[02:13:08](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7988s) measure it rigorously before we put a

[02:13:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7990s) number on it.

[02:13:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7992s) And the third every change that agent

[02:13:14](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7994s) writes still meets a human. Fixes land

[02:13:18](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=7998s) only through normal review and

[02:13:20](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=8000s) continuous integration. And most drafts

[02:13:22](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=8002s) don't land. That's the point. Review is

[02:13:25](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=8005s) the filter.

[02:13:27](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=8007s) Uncle stops being attacks and becomes a

[02:13:29](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=8009s) review cube. Here's

[02:13:32](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=8012s) what we're focusing on next. First,

[02:13:35](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=8015s) mitigations that get approved, not just

[02:13:37](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=8017s) drafted. We great every mitigation that

[02:13:40](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=8020s) agent proposes and feed the outcome back

[02:13:42](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=8022s) so it earns its next level instead of

[02:13:44](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=8024s) being handed it. Second, skills are

[02:13:47](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=8027s) written from the chat so a team can turn

[02:13:50](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=8030s) its runbooks into skills without writing

[02:13:52](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=8032s) any code. And third, measuring the time

[02:13:55](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=8035s) we give back properly with a rigorous

[02:13:57](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=8037s) study of uncall time saved rather than

[02:13:59](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=8039s) usage counts.

[02:14:01](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=8041s) And the next rung on the ladder is uh

[02:14:04](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=8044s) earning medium criticality autonomy with

[02:14:07](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=8047s) evidence. On that, thank you so much for

[02:14:10](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=8050s) listening to us and hopefully that was

[02:14:12](https://www.youtube.com/watch?v=ZZ19zhYnDKY&t=8052s) helpful. Bye.