PRB-4: one real crash, start to finish
Aient case film, with Henrik, founder of Aient

One real crash, start to finish

Hello, I wanted to show you what it looks like when you have Aient installed in an application. Here we have an application called Luffarschack, or Five in a Row. And the general idea is that you try to win against a computer. I'm going to play the computer here. And I'm X, and I want to get five in a row here. So I have to block the computer, and the computer has to block me.

So I'm going to let it win by placing some bad moves here. And now you can see that it is already winning there. And play again, and we have a bug. Oh no, the game is not playable right now. Reduce of empty array with no initial value.

Detect

So then what happens is that this is traced into Aient. And we also get the logs. And these are then taken into the problem coalescing pipeline, which is another way of saying this might happen many times, but it's the same problem. So you don't want to create many problems. Problem is kind of like an issue tracker in Aient. It also synchronizes with Linear. And then when you go to the dashboard of Aient, you can see that it needs attention right now.

Context

And let's go into it. We can see that computerPlayer.ts reduce() called. So this is a title described by a smarter agent. The likely cause here is what is immediately triaged. So the suggested fix is also the later agent that actually looks into it.

Fix

So what the agent has done, what the agent has done is that it's created a pull request, a potential fix here. And it's pushed it to the repository. So we'll have a little look at what it is. We can see that we have the Aient bot generated. It describes the root cause and the change and the validation that it did.

And that it actually tested it. It's passing the checks. And also Aient will be looking at the checks to make sure that everything is, well, checks out. And then when it's done, we can actually just tell it to merge the PR. And then the PR will merge. And the next step is that Aient will verify the PR. So the next time that we play this, when we play the computer and we get a new game, then we shouldn't get the error. It's going to have to build out via Vercel before that happens. So we'll do a little pause and then come back once Vercel has deployed it.

we should be able to see that it's queued here. So let's fast forward a little bit.

Verify

Actually, I figure while it's building, let me show you some features so you can see that it's verifying. You can see that it's verifying and that's because it looks at the SHA commit of this fix or the merge commit of this fix and then it's going to try to correlate that back to the actual crashing location in the codebase. But first it has to receive telemetry using this new SHA that this creates. Here we can see the status. It's monitoring production because we're waiting for the fix to deploy. And the log here is like note taking app for the agent. So once it works through the problem, it adds notes here, which means that your agent can take over if it wants to. Just point it to this problem name or the problem ID up here and it'll pull it by the MCP. Notes is your way of communicating with the agent as on the problem. But you could also go into the thread here, which is like what you'd have in Codex or Claude Code, which is like the streaming work that the agent is performing. You can also show a lot more details here. Every detail we have about the problem, like the category. All the background information that you can query using the API. A little further down, you can see the stack trace exemplar and you can see that it's resolved fully and show you the code where everything happened. And you can finally see example logs. And so when it happened, that's also available here. By going to telemetry, you can see when it happened. So now let's see if our deployment is ready and if we can test again. Okay, so it's been deployed two minutes ago. And let's try it out. Again, playing the computer.

Computer wins. And it worked. And it worked. Amazing. Now what we expect in Aient is that it should have monitored the logs. And it will now know that we don't have a problem anymore.

All right, so now let's have a look at what happens after waiting a little bit. More metrics have come in. We can again notice the release attribution being… using the Git commit SHA. And we see that it has transitioned to Resolved and Verified. Verified is important because now we have evidence that this is not a bug anymore. Additionally, what we could do if we wanted to see more of it is that we could scroll down, and we could see the PR here, we can view on GitHub, and we can see the stack trace. And that closes it out. Thank you for watching.