# Agents Made Code Free. Maintaining It Didn't Get Cheaper.

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

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

[00:00:00](https://www.youtube.com/watch?v=oeknLUsN_GE&t=0s) Hi, I'm Dimmitri. In this session, I

[00:00:03](https://www.youtube.com/watch?v=oeknLUsN_GE&t=3s) will show how I work on maintaining

[00:00:05](https://www.youtube.com/watch?v=oeknLUsN_GE&t=5s) large AI coding projects. Coding agents

[00:00:08](https://www.youtube.com/watch?v=oeknLUsN_GE&t=8s) made code cheap. They made entropy even

[00:00:12](https://www.youtube.com/watch?v=oeknLUsN_GE&t=12s) cheaper. In the Marvel movie, Thanos

[00:00:15](https://www.youtube.com/watch?v=oeknLUsN_GE&t=15s) watched his home planet Titan collapse

[00:00:18](https://www.youtube.com/watch?v=oeknLUsN_GE&t=18s) under its own growth. So, when the

[00:00:20](https://www.youtube.com/watch?v=oeknLUsN_GE&t=20s) universe got crowded, he raised half of

[00:00:23](https://www.youtube.com/watch?v=oeknLUsN_GE&t=23s) everything. This talk is about the one

[00:00:26](https://www.youtube.com/watch?v=oeknLUsN_GE&t=26s) time I did the same in one project. Our

[00:00:30](https://www.youtube.com/watch?v=oeknLUsN_GE&t=30s) code was growing the way Titan did. When

[00:00:32](https://www.youtube.com/watch?v=oeknLUsN_GE&t=32s) I found what my agents had done, I

[00:00:36](https://www.youtube.com/watch?v=oeknLUsN_GE&t=36s) deleted tests and boil the plate code

[00:00:40](https://www.youtube.com/watch?v=oeknLUsN_GE&t=40s) nobody used. I had to rewrite the whole

[00:00:43](https://www.youtube.com/watch?v=oeknLUsN_GE&t=43s) project and I lost some useful pieces.

[00:00:46](https://www.youtube.com/watch?v=oeknLUsN_GE&t=46s) And that's the dilemma. Drown and chaos

[00:00:49](https://www.youtube.com/watch?v=oeknLUsN_GE&t=49s) or break what people use. And here we'll

[00:00:53](https://www.youtube.com/watch?v=oeknLUsN_GE&t=53s) build the third option for that project.

[00:00:56](https://www.youtube.com/watch?v=oeknLUsN_GE&t=56s) The rewrite didn't hold. The agents kept

[00:01:00](https://www.youtube.com/watch?v=oeknLUsN_GE&t=60s) writing and the mess grew back. So I

[00:01:03](https://www.youtube.com/watch?v=oeknLUsN_GE&t=63s) drew a line above it. There is more

[00:01:06](https://www.youtube.com/watch?v=oeknLUsN_GE&t=66s) change than anyone can check. The third

[00:01:09](https://www.youtube.com/watch?v=oeknLUsN_GE&t=69s) option is to stay under it with small

[00:01:12](https://www.youtube.com/watch?v=oeknLUsN_GE&t=72s) sweeps every week and feedback loops

[00:01:15](https://www.youtube.com/watch?v=oeknLUsN_GE&t=75s) built into the system. Each chapter of

[00:01:17](https://www.youtube.com/watch?v=oeknLUsN_GE&t=77s) this talk is one thing I had to learn to

[00:01:19](https://www.youtube.com/watch?v=oeknLUsN_GE&t=79s) get there. Thanos needed six stones for

[00:01:22](https://www.youtube.com/watch?v=oeknLUsN_GE&t=82s) one snap. Ours are six questions and

[00:01:26](https://www.youtube.com/watch?v=oeknLUsN_GE&t=86s) they light up in the corner as we go.

[00:01:29](https://www.youtube.com/watch?v=oeknLUsN_GE&t=89s) The first one is time. How do we stay

[00:01:32](https://www.youtube.com/watch?v=oeknLUsN_GE&t=92s) under the line?

[00:01:34](https://www.youtube.com/watch?v=oeknLUsN_GE&t=94s) Chapter one. To stay under the line. I

[00:01:37](https://www.youtube.com/watch?v=oeknLUsN_GE&t=97s) first had to know what's right means

[00:01:39](https://www.youtube.com/watch?v=oeknLUsN_GE&t=99s) here. I used to think that tests decide

[00:01:42](https://www.youtube.com/watch?v=oeknLUsN_GE&t=102s) that. So what does right actually depend

[00:01:46](https://www.youtube.com/watch?v=oeknLUsN_GE&t=106s) on?

[00:01:48](https://www.youtube.com/watch?v=oeknLUsN_GE&t=108s) Let's keep one task in mind. Someone

[00:01:50](https://www.youtube.com/watch?v=oeknLUsN_GE&t=110s) wants to change the amount on an order

[00:01:53](https://www.youtube.com/watch?v=oeknLUsN_GE&t=113s) we already approved. Two kinds of agents

[00:01:56](https://www.youtube.com/watch?v=oeknLUsN_GE&t=116s) meet here. One wrote the instructions

[00:01:59](https://www.youtube.com/watch?v=oeknLUsN_GE&t=119s) and the other follows them in

[00:02:01](https://www.youtube.com/watch?v=oeknLUsN_GE&t=121s) production.

[00:02:04](https://www.youtube.com/watch?v=oeknLUsN_GE&t=124s) We run this in 110 countries with about

[00:02:07](https://www.youtube.com/watch?v=oeknLUsN_GE&t=127s) 10,000 scenarios of pros instructions.

[00:02:11](https://www.youtube.com/watch?v=oeknLUsN_GE&t=131s) Every task lands in a slice. One process

[00:02:14](https://www.youtube.com/watch?v=oeknLUsN_GE&t=134s) in one country. There's a second stone

[00:02:18](https://www.youtube.com/watch?v=oeknLUsN_GE&t=138s) space. It asks where a change is allowed

[00:02:21](https://www.youtube.com/watch?v=oeknLUsN_GE&t=141s) to go. By default, it stays inside its

[00:02:25](https://www.youtube.com/watch?v=oeknLUsN_GE&t=145s) slice.

[00:02:27](https://www.youtube.com/watch?v=oeknLUsN_GE&t=147s) Most of this system is pros about three

[00:02:29](https://www.youtube.com/watch?v=oeknLUsN_GE&t=149s) quarters by my estimate. So pros needs

[00:02:32](https://www.youtube.com/watch?v=oeknLUsN_GE&t=152s) engineering too. At start we borrowed

[00:02:35](https://www.youtube.com/watch?v=oeknLUsN_GE&t=155s) mutation testing. You break one sentence

[00:02:37](https://www.youtube.com/watch?v=oeknLUsN_GE&t=157s) on purpose and you watch for red test.

[00:02:41](https://www.youtube.com/watch?v=oeknLUsN_GE&t=161s) Here almost none turned red. And if your

[00:02:44](https://www.youtube.com/watch?v=oeknLUsN_GE&t=164s) wrapper is all code, your agent

[00:02:47](https://www.youtube.com/watch?v=oeknLUsN_GE&t=167s) instructions and your tests are your

[00:02:49](https://www.youtube.com/watch?v=oeknLUsN_GE&t=169s) pros. With agents, a rewrite is easy.

[00:02:54](https://www.youtube.com/watch?v=oeknLUsN_GE&t=174s) The hard part is knowing what has to

[00:02:56](https://www.youtube.com/watch?v=oeknLUsN_GE&t=176s) survive it. For code, tests can tell

[00:02:59](https://www.youtube.com/watch?v=oeknLUsN_GE&t=179s) you. For pros, only people can. Agents

[00:03:02](https://www.youtube.com/watch?v=oeknLUsN_GE&t=182s) can't build a complex system on their

[00:03:05](https://www.youtube.com/watch?v=oeknLUsN_GE&t=185s) own. They fix what is easy to see and

[00:03:08](https://www.youtube.com/watch?v=oeknLUsN_GE&t=188s) break consistency on the way. And people

[00:03:11](https://www.youtube.com/watch?v=oeknLUsN_GE&t=191s) also get tired. So we spend their

[00:03:13](https://www.youtube.com/watch?v=oeknLUsN_GE&t=193s) judgment only where it's needed. When

[00:03:17](https://www.youtube.com/watch?v=oeknLUsN_GE&t=197s) business asked check with the team, it

[00:03:20](https://www.youtube.com/watch?v=oeknLUsN_GE&t=200s) hides a real question. Is this a human

[00:03:23](https://www.youtube.com/watch?v=oeknLUsN_GE&t=203s) step or can we automate it? Experts

[00:03:25](https://www.youtube.com/watch?v=oeknLUsN_GE&t=205s) can't answer that. They don't see how

[00:03:28](https://www.youtube.com/watch?v=oeknLUsN_GE&t=208s) the systems connect inside. So

[00:03:30](https://www.youtube.com/watch?v=oeknLUsN_GE&t=210s) developers dig through the traces and

[00:03:33](https://www.youtube.com/watch?v=oeknLUsN_GE&t=213s) logs with their coding agents and they

[00:03:35](https://www.youtube.com/watch?v=oeknLUsN_GE&t=215s) decide which it is. In our task, the

[00:03:38](https://www.youtube.com/watch?v=oeknLUsN_GE&t=218s) team leads decision stays human. The

[00:03:41](https://www.youtube.com/watch?v=oeknLUsN_GE&t=221s) limit was data we could connect.

[00:03:45](https://www.youtube.com/watch?v=oeknLUsN_GE&t=225s) Business experts see the rules.

[00:03:47](https://www.youtube.com/watch?v=oeknLUsN_GE&t=227s) Developers see what's under the water.

[00:03:50](https://www.youtube.com/watch?v=oeknLUsN_GE&t=230s) And that's where most bugs leave. They

[00:03:53](https://www.youtube.com/watch?v=oeknLUsN_GE&t=233s) have their coding agents write those

[00:03:55](https://www.youtube.com/watch?v=oeknLUsN_GE&t=235s) links down so the runtime can check

[00:03:58](https://www.youtube.com/watch?v=oeknLUsN_GE&t=238s) them. Our guide for experts says it

[00:04:00](https://www.youtube.com/watch?v=oeknLUsN_GE&t=240s) simply business tell us what we work out

[00:04:04](https://www.youtube.com/watch?v=oeknLUsN_GE&t=244s) how. One document to readers. Agents

[00:04:09](https://www.youtube.com/watch?v=oeknLUsN_GE&t=249s) read the blogs. So their blocks have to

[00:04:11](https://www.youtube.com/watch?v=oeknLUsN_GE&t=251s) be exact. Experts read the text around

[00:04:14](https://www.youtube.com/watch?v=oeknLUsN_GE&t=254s) them and they sign it. We keep one copy,

[00:04:17](https://www.youtube.com/watch?v=oeknLUsN_GE&t=257s) a second one with drift and works both

[00:04:20](https://www.youtube.com/watch?v=oeknLUsN_GE&t=260s) ways. Our code needs a plain description

[00:04:22](https://www.youtube.com/watch?v=oeknLUsN_GE&t=262s) for experts too. Experts can only fix

[00:04:26](https://www.youtube.com/watch?v=oeknLUsN_GE&t=266s) what they can see. Each of them reviews

[00:04:29](https://www.youtube.com/watch?v=oeknLUsN_GE&t=269s) a task exactly the way the agent saw it.

[00:04:33](https://www.youtube.com/watch?v=oeknLUsN_GE&t=273s) The node becomes the next change made by

[00:04:35](https://www.youtube.com/watch?v=oeknLUsN_GE&t=275s) a coding agent and test it on the same

[00:04:38](https://www.youtube.com/watch?v=oeknLUsN_GE&t=278s) task. So right depends on people. No

[00:04:42](https://www.youtube.com/watch?v=oeknLUsN_GE&t=282s) single person knows it. A group of

[00:04:45](https://www.youtube.com/watch?v=oeknLUsN_GE&t=285s) experts decides and that's why we bring

[00:04:48](https://www.youtube.com/watch?v=oeknLUsN_GE&t=288s) them in from the start. You'll see the

[00:04:51](https://www.youtube.com/watch?v=oeknLUsN_GE&t=291s) soul stone here a lot. We take it in

[00:04:54](https://www.youtube.com/watch?v=oeknLUsN_GE&t=294s) chapter 4.

[00:04:56](https://www.youtube.com/watch?v=oeknLUsN_GE&t=296s) Chapter two. Experts correct one task at

[00:04:59](https://www.youtube.com/watch?v=oeknLUsN_GE&t=299s) a time while agents run the same pros

[00:05:02](https://www.youtube.com/watch?v=oeknLUsN_GE&t=302s) all day. We track every outcome accepted

[00:05:06](https://www.youtube.com/watch?v=oeknLUsN_GE&t=306s) or rejected. The accuracy shows us where

[00:05:10](https://www.youtube.com/watch?v=oeknLUsN_GE&t=310s) people still need to look. I used to

[00:05:13](https://www.youtube.com/watch?v=oeknLUsN_GE&t=313s) think the goal was to move everything

[00:05:15](https://www.youtube.com/watch?v=oeknLUsN_GE&t=315s) into code. So where do we still need

[00:05:18](https://www.youtube.com/watch?v=oeknLUsN_GE&t=318s) judgment? Here is our task again. The

[00:05:21](https://www.youtube.com/watch?v=oeknLUsN_GE&t=321s) pro says the limit is 1,500. The code

[00:05:25](https://www.youtube.com/watch?v=oeknLUsN_GE&t=325s) says a,000. The same order is fine in

[00:05:28](https://www.youtube.com/watch?v=oeknLUsN_GE&t=328s) one place and too high in the other and

[00:05:31](https://www.youtube.com/watch?v=oeknLUsN_GE&t=331s) nothing throws an error. Even a line

[00:05:34](https://www.youtube.com/watch?v=oeknLUsN_GE&t=334s) like do not call X teaches the model

[00:05:37](https://www.youtube.com/watch?v=oeknLUsN_GE&t=337s) that X exists. That's the third stone

[00:05:41](https://www.youtube.com/watch?v=oeknLUsN_GE&t=341s) reality. Is it true or does it only look

[00:05:44](https://www.youtube.com/watch?v=oeknLUsN_GE&t=344s) true for the model? Pros has no

[00:05:47](https://www.youtube.com/watch?v=oeknLUsN_GE&t=347s) compiler. So we built one. Then the

[00:05:50](https://www.youtube.com/watch?v=oeknLUsN_GE&t=350s) author lints every changed file into

[00:05:53](https://www.youtube.com/watch?v=oeknLUsN_GE&t=353s) ledger just like code.

[00:05:58](https://www.youtube.com/watch?v=oeknLUsN_GE&t=358s) Lind can't stop copies drifting aparts.

[00:06:02](https://www.youtube.com/watch?v=oeknLUsN_GE&t=362s) So we watched the walks. A walk is the

[00:06:04](https://www.youtube.com/watch?v=oeknLUsN_GE&t=364s) path one agent takes through the pros.

[00:06:07](https://www.youtube.com/watch?v=oeknLUsN_GE&t=367s) When agents keep walking in the same

[00:06:09](https://www.youtube.com/watch?v=oeknLUsN_GE&t=369s) path, we turn it into one workflow tool

[00:06:12](https://www.youtube.com/watch?v=oeknLUsN_GE&t=372s) and we delete the pros that walked it.

[00:06:16](https://www.youtube.com/watch?v=oeknLUsN_GE&t=376s) Our first workflow of tools made things

[00:06:19](https://www.youtube.com/watch?v=oeknLUsN_GE&t=379s) worse. The pros still named the steps

[00:06:22](https://www.youtube.com/watch?v=oeknLUsN_GE&t=382s) inside them. So agents run those steps

[00:06:25](https://www.youtube.com/watch?v=oeknLUsN_GE&t=385s) again. It worked once. The pros kept

[00:06:27](https://www.youtube.com/watch?v=oeknLUsN_GE&t=387s) only the decisions. Now a walk becomes

[00:06:31](https://www.youtube.com/watch?v=oeknLUsN_GE&t=391s) code only when it repeats and agrees

[00:06:34](https://www.youtube.com/watch?v=oeknLUsN_GE&t=394s) across countries otherwise it stays

[00:06:37](https://www.youtube.com/watch?v=oeknLUsN_GE&t=397s) pros. What still varies stays in pros

[00:06:41](https://www.youtube.com/watch?v=oeknLUsN_GE&t=401s) inside the frame. Code prepares the

[00:06:44](https://www.youtube.com/watch?v=oeknLUsN_GE&t=404s) facts and carries out the decision.

[00:06:47](https://www.youtube.com/watch?v=oeknLUsN_GE&t=407s) Experts own the decision table and the

[00:06:50](https://www.youtube.com/watch?v=oeknLUsN_GE&t=410s) model only picks a row. If no row fits,

[00:06:54](https://www.youtube.com/watch?v=oeknLUsN_GE&t=414s) a person takes over. We haven't built

[00:06:56](https://www.youtube.com/watch?v=oeknLUsN_GE&t=416s) the next step yet. The model will take

[00:06:58](https://www.youtube.com/watch?v=oeknLUsN_GE&t=418s) the small decisions and workflows will

[00:07:01](https://www.youtube.com/watch?v=oeknLUsN_GE&t=421s) connect them. This boundary moves both

[00:07:04](https://www.youtube.com/watch?v=oeknLUsN_GE&t=424s) ways. We moved hundreds of lines of

[00:07:07](https://www.youtube.com/watch?v=oeknLUsN_GE&t=427s) expressions out of code and into pros

[00:07:11](https://www.youtube.com/watch?v=oeknLUsN_GE&t=431s) because experts couldn't read them. A

[00:07:14](https://www.youtube.com/watch?v=oeknLUsN_GE&t=434s) reax is still fine for spotting a

[00:07:15](https://www.youtube.com/watch?v=oeknLUsN_GE&t=435s) critical signal. When there are many,

[00:07:18](https://www.youtube.com/watch?v=oeknLUsN_GE&t=438s) they move into config. Judgment stays

[00:07:21](https://www.youtube.com/watch?v=oeknLUsN_GE&t=441s) wherever the answer varies in a place an

[00:07:24](https://www.youtube.com/watch?v=oeknLUsN_GE&t=444s) expert can read. Chapter three. Now the

[00:07:28](https://www.youtube.com/watch?v=oeknLUsN_GE&t=448s) model decides and its decision becomes

[00:07:31](https://www.youtube.com/watch?v=oeknLUsN_GE&t=451s) an action. I used to think control meant

[00:07:35](https://www.youtube.com/watch?v=oeknLUsN_GE&t=455s) permissions on the agent. So what does

[00:07:37](https://www.youtube.com/watch?v=oeknLUsN_GE&t=457s) it mean to stay in control?

[00:07:41](https://www.youtube.com/watch?v=oeknLUsN_GE&t=461s) Control starts with what the model sees.

[00:07:44](https://www.youtube.com/watch?v=oeknLUsN_GE&t=464s) That's what the fourth stone mind. What

[00:07:47](https://www.youtube.com/watch?v=oeknLUsN_GE&t=467s) does the model know? When we let it

[00:07:49](https://www.youtube.com/watch?v=oeknLUsN_GE&t=469s) choose what to load, it loaded

[00:07:52](https://www.youtube.com/watch?v=oeknLUsN_GE&t=472s) everything. Then it invented a pricing

[00:07:55](https://www.youtube.com/watch?v=oeknLUsN_GE&t=475s) check just because prices were visible.

[00:07:58](https://www.youtube.com/watch?v=oeknLUsN_GE&t=478s) Whatever sits in its context becomes a

[00:08:01](https://www.youtube.com/watch?v=oeknLUsN_GE&t=481s) reason to act. The runtime decides

[00:08:04](https://www.youtube.com/watch?v=oeknLUsN_GE&t=484s) what's the model sees for our task.

[00:08:07](https://www.youtube.com/watch?v=oeknLUsN_GE&t=487s) That's three facts from the order. The

[00:08:09](https://www.youtube.com/watch?v=oeknLUsN_GE&t=489s) full record stays with the runtime and

[00:08:12](https://www.youtube.com/watch?v=oeknLUsN_GE&t=492s) the model makes fewer wrong moves.

[00:08:17](https://www.youtube.com/watch?v=oeknLUsN_GE&t=497s) Here how we minimize context blo for a

[00:08:20](https://www.youtube.com/watch?v=oeknLUsN_GE&t=500s) big set. The model passes a handle, a

[00:08:22](https://www.youtube.com/watch?v=oeknLUsN_GE&t=502s) short name for date that the runtime

[00:08:24](https://www.youtube.com/watch?v=oeknLUsN_GE&t=504s) already holds. The runtime checks it

[00:08:27](https://www.youtube.com/watch?v=oeknLUsN_GE&t=507s) before any write. It keeps the context

[00:08:29](https://www.youtube.com/watch?v=oeknLUsN_GE&t=509s) small and still enough for the decision.

[00:08:32](https://www.youtube.com/watch?v=oeknLUsN_GE&t=512s) At the same time, it uh validates

[00:08:35](https://www.youtube.com/watch?v=oeknLUsN_GE&t=515s) context for consistency. A model with

[00:08:39](https://www.youtube.com/watch?v=oeknLUsN_GE&t=519s) the right facts can still do damage.

[00:08:41](https://www.youtube.com/watch?v=oeknLUsN_GE&t=521s) That's the fifth stone, power. What made

[00:08:45](https://www.youtube.com/watch?v=oeknLUsN_GE&t=525s) change? We control a change twice.

[00:08:47](https://www.youtube.com/watch?v=oeknLUsN_GE&t=527s) during development and again at the

[00:08:50](https://www.youtube.com/watch?v=oeknLUsN_GE&t=530s) right at runtime only code opens the

[00:08:53](https://www.youtube.com/watch?v=oeknLUsN_GE&t=533s) right path so we can change pros many

[00:08:56](https://www.youtube.com/watch?v=oeknLUsN_GE&t=536s) times a day the open rings are the parts

[00:08:59](https://www.youtube.com/watch?v=oeknLUsN_GE&t=539s) that still depend on people let's zoom

[00:09:02](https://www.youtube.com/watch?v=oeknLUsN_GE&t=542s) in and run right five walls stand in

[00:09:06](https://www.youtube.com/watch?v=oeknLUsN_GE&t=546s) front of it and the sixth checks that it

[00:09:08](https://www.youtube.com/watch?v=oeknLUsN_GE&t=548s) landed if a right looks like it failed

[00:09:12](https://www.youtube.com/watch?v=oeknLUsN_GE&t=552s) an agent retries and the journal makes

[00:09:15](https://www.youtube.com/watch?v=oeknLUsN_GE&t=555s) sure it runs only once means much more

[00:09:18](https://www.youtube.com/watch?v=oeknLUsN_GE&t=558s) than permissions on the agent.

[00:09:24](https://www.youtube.com/watch?v=oeknLUsN_GE&t=564s) Chapter 4. Most days I glide over the

[00:09:27](https://www.youtube.com/watch?v=oeknLUsN_GE&t=567s) system. From up here, everything looks

[00:09:30](https://www.youtube.com/watch?v=oeknLUsN_GE&t=570s) fine, so I don't go down into every

[00:09:32](https://www.youtube.com/watch?v=oeknLUsN_GE&t=572s) change my agents made. Then the surface

[00:09:36](https://www.youtube.com/watch?v=oeknLUsN_GE&t=576s) gets rough. How do we know we're losing

[00:09:39](https://www.youtube.com/watch?v=oeknLUsN_GE&t=579s) control?

[00:09:41](https://www.youtube.com/watch?v=oeknLUsN_GE&t=581s) All those controls protect the system.

[00:09:43](https://www.youtube.com/watch?v=oeknLUsN_GE&t=583s) Nothing protected my attention. At the

[00:09:46](https://www.youtube.com/watch?v=oeknLUsN_GE&t=586s) start, I read every change and that's

[00:09:48](https://www.youtube.com/watch?v=oeknLUsN_GE&t=588s) how the system picked up my habits. Then

[00:09:52](https://www.youtube.com/watch?v=oeknLUsN_GE&t=592s) the volume grew past what one person can

[00:09:55](https://www.youtube.com/watch?v=oeknLUsN_GE&t=595s) read. I was burning out and I stopped

[00:09:58](https://www.youtube.com/watch?v=oeknLUsN_GE&t=598s) reading every change. That was the price

[00:10:00](https://www.youtube.com/watch?v=oeknLUsN_GE&t=600s) of the last stone. So, who judges and

[00:10:04](https://www.youtube.com/watch?v=oeknLUsN_GE&t=604s) what does it cost? With it, the gauntlet

[00:10:07](https://www.youtube.com/watch?v=oeknLUsN_GE&t=607s) is complete. Every tech lead I know has

[00:10:10](https://www.youtube.com/watch?v=oeknLUsN_GE&t=610s) been here. The one person every change

[00:10:13](https://www.youtube.com/watch?v=oeknLUsN_GE&t=613s) waits on. agents just get you there in

[00:10:16](https://www.youtube.com/watch?v=oeknLUsN_GE&t=616s) weeks instead of years. I needed a way

[00:10:19](https://www.youtube.com/watch?v=oeknLUsN_GE&t=619s) of working that doesn't run through one

[00:10:22](https://www.youtube.com/watch?v=oeknLUsN_GE&t=622s) head. And the first gauge was me. When

[00:10:25](https://www.youtube.com/watch?v=oeknLUsN_GE&t=625s) I'm tired and I can't explain what

[00:10:27](https://www.youtube.com/watch?v=oeknLUsN_GE&t=627s) changed, that's the signal. The

[00:10:30](https://www.youtube.com/watch?v=oeknLUsN_GE&t=630s) dashboard can be green while that gauge

[00:10:33](https://www.youtube.com/watch?v=oeknLUsN_GE&t=633s) is firing. So, I stop and fix the

[00:10:36](https://www.youtube.com/watch?v=oeknLUsN_GE&t=636s) process first. I don't wait for a

[00:10:38](https://www.youtube.com/watch?v=oeknLUsN_GE&t=638s) number. I couldn't read every div

[00:10:41](https://www.youtube.com/watch?v=oeknLUsN_GE&t=641s) anymore. So I built instruments that

[00:10:44](https://www.youtube.com/watch?v=oeknLUsN_GE&t=644s) watch how things change and their

[00:10:46](https://www.youtube.com/watch?v=oeknLUsN_GE&t=646s) derivatives. One of them is blast

[00:10:49](https://www.youtube.com/watch?v=oeknLUsN_GE&t=649s) radius. Before we built one new check,

[00:10:52](https://www.youtube.com/watch?v=oeknLUsN_GE&t=652s) we replayed its rule over several

[00:10:54](https://www.youtube.com/watch?v=oeknLUsN_GE&t=654s) thousand past sessions. It would have

[00:10:56](https://www.youtube.com/watch?v=oeknLUsN_GE&t=656s) fired on less than 1%. Where the numbers

[00:11:00](https://www.youtube.com/watch?v=oeknLUsN_GE&t=660s) jump, I go down and read. When something

[00:11:04](https://www.youtube.com/watch?v=oeknLUsN_GE&t=664s) looks wrong, we wait. One signal waits

[00:11:06](https://www.youtube.com/watch?v=oeknLUsN_GE&t=666s) until it comes back, even when it comes

[00:11:09](https://www.youtube.com/watch?v=oeknLUsN_GE&t=669s) from an expert. A cluster across slices

[00:11:12](https://www.youtube.com/watch?v=oeknLUsN_GE&t=672s) is different. Then we fix the concept

[00:11:15](https://www.youtube.com/watch?v=oeknLUsN_GE&t=675s) behind it once. Left alone, an agent

[00:11:18](https://www.youtube.com/watch?v=oeknLUsN_GE&t=678s) keeps adding conditions until today's

[00:11:20](https://www.youtube.com/watch?v=oeknLUsN_GE&t=680s) tasks pass. The eval score drops when

[00:11:23](https://www.youtube.com/watch?v=oeknLUsN_GE&t=683s) you take them out and you get an honest

[00:11:26](https://www.youtube.com/watch?v=oeknLUsN_GE&t=686s) list of gaps.

[00:11:31](https://www.youtube.com/watch?v=oeknLUsN_GE&t=691s) One red slice is rarely alone. The

[00:11:34](https://www.youtube.com/watch?v=oeknLUsN_GE&t=694s) experts note becomes an eval and an

[00:11:36](https://www.youtube.com/watch?v=oeknLUsN_GE&t=696s) agent searches the siblings for the same

[00:11:39](https://www.youtube.com/watch?v=oeknLUsN_GE&t=699s) failure. Siblings differ on purpose by

[00:11:42](https://www.youtube.com/watch?v=oeknLUsN_GE&t=702s) market and by local law. So we sweep

[00:11:45](https://www.youtube.com/watch?v=oeknLUsN_GE&t=705s) them and the fix goes only where the

[00:11:48](https://www.youtube.com/watch?v=oeknLUsN_GE&t=708s) failure is confirmed.

[00:11:53](https://www.youtube.com/watch?v=oeknLUsN_GE&t=713s) At any time 10 to 20 of these loops are

[00:11:57](https://www.youtube.com/watch?v=oeknLUsN_GE&t=717s) open. At one point I had become the

[00:11:59](https://www.youtube.com/watch?v=oeknLUsN_GE&t=719s) message bus between users and my agents.

[00:12:03](https://www.youtube.com/watch?v=oeknLUsN_GE&t=723s) At this scale, every user has to be an

[00:12:06](https://www.youtube.com/watch?v=oeknLUsN_GE&t=726s) evaluator with a screen to give feedback

[00:12:08](https://www.youtube.com/watch?v=oeknLUsN_GE&t=728s) without me. My job as the architect is

[00:12:12](https://www.youtube.com/watch?v=oeknLUsN_GE&t=732s) to build those loops into the system.

[00:12:15](https://www.youtube.com/watch?v=oeknLUsN_GE&t=735s) Each loop writes to a ledger. So nothing

[00:12:18](https://www.youtube.com/watch?v=oeknLUsN_GE&t=738s) waits on my memory and nothing waits on

[00:12:20](https://www.youtube.com/watch?v=oeknLUsN_GE&t=740s) me.

[00:12:24](https://www.youtube.com/watch?v=oeknLUsN_GE&t=744s) Agents forget everything when a session

[00:12:26](https://www.youtube.com/watch?v=oeknLUsN_GE&t=746s) ends. So the memory has to leave in

[00:12:29](https://www.youtube.com/watch?v=oeknLUsN_GE&t=749s) files. We keep two kinds. what happened

[00:12:32](https://www.youtube.com/watch?v=oeknLUsN_GE&t=752s) and what is true now. The next session

[00:12:35](https://www.youtube.com/watch?v=oeknLUsN_GE&t=755s) starts from what is true now.

[00:12:40](https://www.youtube.com/watch?v=oeknLUsN_GE&t=760s) Every issue has to connect to its root

[00:12:43](https://www.youtube.com/watch?v=oeknLUsN_GE&t=763s) cause. Without that, the agent decides

[00:12:46](https://www.youtube.com/watch?v=oeknLUsN_GE&t=766s) on wrong facts and the check has nothing

[00:12:49](https://www.youtube.com/watch?v=oeknLUsN_GE&t=769s) to count. So we don't act without

[00:12:51](https://www.youtube.com/watch?v=oeknLUsN_GE&t=771s) evidence.

[00:12:54](https://www.youtube.com/watch?v=oeknLUsN_GE&t=774s) Reality comes back. In chapter two, it

[00:12:58](https://www.youtube.com/watch?v=oeknLUsN_GE&t=778s) asked if the pros is true. Now it asks

[00:13:01](https://www.youtube.com/watch?v=oeknLUsN_GE&t=781s) if the change is I used to trust green.

[00:13:05](https://www.youtube.com/watch?v=oeknLUsN_GE&t=785s) So what would make us change our mind?

[00:13:08](https://www.youtube.com/watch?v=oeknLUsN_GE&t=788s) Green was tempting because checking is

[00:13:11](https://www.youtube.com/watch?v=oeknLUsN_GE&t=791s) where the times goes. My agent spend

[00:13:14](https://www.youtube.com/watch?v=oeknLUsN_GE&t=794s) almost 9/10en of that time on a change

[00:13:17](https://www.youtube.com/watch?v=oeknLUsN_GE&t=797s) just checking it and checking took a

[00:13:19](https://www.youtube.com/watch?v=oeknLUsN_GE&t=799s) bigger share of my own week too.

[00:13:22](https://www.youtube.com/watch?v=oeknLUsN_GE&t=802s) Verification speed is the limit.

[00:13:27](https://www.youtube.com/watch?v=oeknLUsN_GE&t=807s) We rank evidence by what it proves. A

[00:13:30](https://www.youtube.com/watch?v=oeknLUsN_GE&t=810s) replay on real tasks sits at the top.

[00:13:34](https://www.youtube.com/watch?v=oeknLUsN_GE&t=814s) Before the data arrives, we write down

[00:13:36](https://www.youtube.com/watch?v=oeknLUsN_GE&t=816s) what would prove us wrong. And every new

[00:13:39](https://www.youtube.com/watch?v=oeknLUsN_GE&t=819s) check has to fail once on purpose before

[00:13:42](https://www.youtube.com/watch?v=oeknLUsN_GE&t=822s) we trust it.

[00:13:46](https://www.youtube.com/watch?v=oeknLUsN_GE&t=826s) AI refactors fail quietly. Every

[00:13:50](https://www.youtube.com/watch?v=oeknLUsN_GE&t=830s) condition and every decision table row

[00:13:53](https://www.youtube.com/watch?v=oeknLUsN_GE&t=833s) is a branch. And after a refactor, each

[00:13:56](https://www.youtube.com/watch?v=oeknLUsN_GE&t=836s) one has to land exactly once. Here the

[00:13:59](https://www.youtube.com/watch?v=oeknLUsN_GE&t=839s) branch for amounts over the limit got

[00:14:01](https://www.youtube.com/watch?v=oeknLUsN_GE&t=841s) lost. Our [snorts] order would have gone

[00:14:04](https://www.youtube.com/watch?v=oeknLUsN_GE&t=844s) through without the team lead. Real

[00:14:07](https://www.youtube.com/watch?v=oeknLUsN_GE&t=847s) tasks decide what we test. We freeze

[00:14:11](https://www.youtube.com/watch?v=oeknLUsN_GE&t=851s) them as a set and replay the old version

[00:14:14](https://www.youtube.com/watch?v=oeknLUsN_GE&t=854s) against the new one. If no real task

[00:14:17](https://www.youtube.com/watch?v=oeknLUsN_GE&t=857s) takes a branch, we call that no

[00:14:19](https://www.youtube.com/watch?v=oeknLUsN_GE&t=859s) evidence.

[00:14:21](https://www.youtube.com/watch?v=oeknLUsN_GE&t=861s) You can't trust one agent to judge its

[00:14:24](https://www.youtube.com/watch?v=oeknLUsN_GE&t=864s) own work. A second agent checks it with

[00:14:28](https://www.youtube.com/watch?v=oeknLUsN_GE&t=868s) a checklist and without the author's

[00:14:30](https://www.youtube.com/watch?v=oeknLUsN_GE&t=870s) reasoning. It can't be talked into

[00:14:32](https://www.youtube.com/watch?v=oeknLUsN_GE&t=872s) agreeing. Over 6 months, it grew to most

[00:14:36](https://www.youtube.com/watch?v=oeknLUsN_GE&t=876s) of my checking.

[00:14:39](https://www.youtube.com/watch?v=oeknLUsN_GE&t=879s) Coding agents finish most of a batch and

[00:14:43](https://www.youtube.com/watch?v=oeknLUsN_GE&t=883s) quietly skipped the rest, especially

[00:14:45](https://www.youtube.com/watch?v=oeknLUsN_GE&t=885s) when their context is full. The ledger

[00:14:48](https://www.youtube.com/watch?v=oeknLUsN_GE&t=888s) shows what was skipped in one backlog.

[00:14:51](https://www.youtube.com/watch?v=oeknLUsN_GE&t=891s) Most items looks covered, but only about

[00:14:54](https://www.youtube.com/watch?v=oeknLUsN_GE&t=894s) a quarter could really be closed. So now

[00:14:58](https://www.youtube.com/watch?v=oeknLUsN_GE&t=898s) evidence we can count is what changes

[00:15:01](https://www.youtube.com/watch?v=oeknLUsN_GE&t=901s) our mind. And this is the third option

[00:15:05](https://www.youtube.com/watch?v=oeknLUsN_GE&t=905s) from the start of the talk. I used to

[00:15:08](https://www.youtube.com/watch?v=oeknLUsN_GE&t=908s) think there were two choices. Keep

[00:15:11](https://www.youtube.com/watch?v=oeknLUsN_GE&t=911s) everything or snap. So what is worth

[00:15:14](https://www.youtube.com/watch?v=oeknLUsN_GE&t=914s) carrying forward?

[00:15:16](https://www.youtube.com/watch?v=oeknLUsN_GE&t=916s) This one process over a year. Some

[00:15:19](https://www.youtube.com/watch?v=oeknLUsN_GE&t=919s) branches merged back, some still drift,

[00:15:22](https://www.youtube.com/watch?v=oeknLUsN_GE&t=922s) and some are dead but still look alive.

[00:15:25](https://www.youtube.com/watch?v=oeknLUsN_GE&t=925s) Which ones deserve to survive?

[00:15:30](https://www.youtube.com/watch?v=oeknLUsN_GE&t=930s) A fix is only a moment. These forces

[00:15:33](https://www.youtube.com/watch?v=oeknLUsN_GE&t=933s) keep pushing right after it. You've met

[00:15:36](https://www.youtube.com/watch?v=oeknLUsN_GE&t=936s) most of them already. This chapter deals

[00:15:39](https://www.youtube.com/watch?v=oeknLUsN_GE&t=939s) with regrowth and branching.

[00:15:44](https://www.youtube.com/watch?v=oeknLUsN_GE&t=944s) The first gauge was me back in chapter

[00:15:46](https://www.youtube.com/watch?v=oeknLUsN_GE&t=946s) 4. These are the gauges for the code.

[00:15:50](https://www.youtube.com/watch?v=oeknLUsN_GE&t=950s) Time gets its answers here. The bridge

[00:15:53](https://www.youtube.com/watch?v=oeknLUsN_GE&t=953s) line comes back as a number. A ratchet

[00:15:56](https://www.youtube.com/watch?v=oeknLUsN_GE&t=956s) means the count can only go down. Two

[00:15:58](https://www.youtube.com/watch?v=oeknLUsN_GE&t=958s) gauges run today and six more are next.

[00:16:04](https://www.youtube.com/watch?v=oeknLUsN_GE&t=964s) Every week real use decides what stays.

[00:16:07](https://www.youtube.com/watch?v=oeknLUsN_GE&t=967s) Anything no real task reaches gets a

[00:16:10](https://www.youtube.com/watch?v=oeknLUsN_GE&t=970s) name and its evidence before it goes.

[00:16:13](https://www.youtube.com/watch?v=oeknLUsN_GE&t=973s) Test the evidence I didn't have when I

[00:16:16](https://www.youtube.com/watch?v=oeknLUsN_GE&t=976s) snapped. Tests kept that variants alive.

[00:16:20](https://www.youtube.com/watch?v=oeknLUsN_GE&t=980s) Agents kept every option a test still

[00:16:23](https://www.youtube.com/watch?v=oeknLUsN_GE&t=983s) mentioned. Now retired variant takes its

[00:16:27](https://www.youtube.com/watch?v=oeknLUsN_GE&t=987s) test with it. Here's the whole method as

[00:16:30](https://www.youtube.com/watch?v=oeknLUsN_GE&t=990s) one week. This is the third option as a

[00:16:34](https://www.youtube.com/watch?v=oeknLUsN_GE&t=994s) process.

[00:16:39](https://www.youtube.com/watch?v=oeknLUsN_GE&t=999s) together. That's the full gauntlet.

[00:16:42](https://www.youtube.com/watch?v=oeknLUsN_GE&t=1002s) Thanos used his for one snap. We use

[00:16:45](https://www.youtube.com/watch?v=oeknLUsN_GE&t=1005s) ours every week in small doses with

[00:16:48](https://www.youtube.com/watch?v=oeknLUsN_GE&t=1008s) evidence. I haven't had to snap again.

[00:16:55](https://www.youtube.com/watch?v=oeknLUsN_GE&t=1015s) Soul had a second price and I only saw

[00:16:58](https://www.youtube.com/watch?v=oeknLUsN_GE&t=1018s) it later. After we deliver it to

[00:17:01](https://www.youtube.com/watch?v=oeknLUsN_GE&t=1021s) production, I found my own engineering

[00:17:03](https://www.youtube.com/watch?v=oeknLUsN_GE&t=1023s) principles built into the systems

[00:17:06](https://www.youtube.com/watch?v=oeknLUsN_GE&t=1026s) design. Its skills and its memory carry

[00:17:09](https://www.youtube.com/watch?v=oeknLUsN_GE&t=1029s) my habits and my biases. Conway said the

[00:17:12](https://www.youtube.com/watch?v=oeknLUsN_GE&t=1032s) system copies the orc chart. Ours copies

[00:17:16](https://www.youtube.com/watch?v=oeknLUsN_GE&t=1036s) whoever corrects it. Here are 18

[00:17:19](https://www.youtube.com/watch?v=oeknLUsN_GE&t=1039s) questions, three for each stone. I'm

[00:17:22](https://www.youtube.com/watch?v=oeknLUsN_GE&t=1042s) using them when working on my projects.

[00:17:26](https://www.youtube.com/watch?v=oeknLUsN_GE&t=1046s) Pause here and pick one your own

[00:17:29](https://www.youtube.com/watch?v=oeknLUsN_GE&t=1049s) repository can't answer yet. That's

[00:17:32](https://www.youtube.com/watch?v=oeknLUsN_GE&t=1052s) where to start.

[00:17:37](https://www.youtube.com/watch?v=oeknLUsN_GE&t=1057s) I learned to use this process when I was

[00:17:40](https://www.youtube.com/watch?v=oeknLUsN_GE&t=1060s) already burning out. If you're the one

[00:17:43](https://www.youtube.com/watch?v=oeknLUsN_GE&t=1063s) your team's changes weight on today,

[00:17:46](https://www.youtube.com/watch?v=oeknLUsN_GE&t=1066s) this is for you. Use your stones while

[00:17:49](https://www.youtube.com/watch?v=oeknLUsN_GE&t=1069s) you can still read every change. In the

[00:17:52](https://www.youtube.com/watch?v=oeknLUsN_GE&t=1072s) end, Thanos retires to a garden. Maybe

[00:17:55](https://www.youtube.com/watch?v=oeknLUsN_GE&t=1075s) he should start using them daily. Thank

[00:17:59](https://www.youtube.com/watch?v=oeknLUsN_GE&t=1079s) you.