Stop Prompting — Greg Pstrucha, Sentry

0:00 All >> [music] >> right. Hello everyone. I'm going to try to speak as loud as I can so it's reasonably good with all the

0:20 all the commotion. Uh my name is Greg. I'm an AI engineer at Sentry. I work on the infrastructure that powers seir centuries uh debugging agent and I have a little rage baity title of the

0:32 presentation that calls to stop prompting but I do not mean loop engineering. I'm going to get uh a little bit deeper here. Uh who still reads code that AI generates?

0:43 Anyone? It's not a trick question and I'm not trying to shame you. I I think it's contextual. I read code when it's sentry code and I need to make sure that we are

0:54 not breaking the actual production for sentry but at the same time if I am working on a side project it doesn't really matter uh to read every line of code and I can I can be a little bit

1:05 more lenient with it uh but then every now and then I go back to the actual code I want to make improvements and um it's bad it's it's overly defensive code it's tests are crap they are not testing

1:19 anything. Um, I would not want to maintain or work in that codebase anymore. So then what you do is you you talk to the agent. You say, "Hey, you've done a mistake. This is overly

1:31 complicated. The code's too defensive. Um, you don't have to create endpoints for everything you're doing, otherwise I have to maintain it forever." And I had to remove a lot of swearing from this

1:41 slide because in my actual transcripts, there is going to be a lot of a lot of rage. Usually those are very elementary mistakes that the agents are making if you don't put any guard rails in place

1:51 and um and that just outrages me. So then you do that, the agent says, "You're absolutely right. Let me fix this." And it's going to run and try to fix the the mistakes.

2:06 And five minutes pass and compaction of context happens. And we're back to back to square one. And you would have to now repeat yourself, reprompt your fixes, reprompt

2:18 the agent on the right track. And my entire thesis for this talk, my entire claim here is you should stop prompting. You should start codifying the actual rules that the agent should

2:31 be following. um rules and policies that you want to establish inside your codebase that are going to capture both convention and the quality of code that you are expecting and would make you

2:43 proud. And so when I say you would want to add deterministic checks or rules, what are you thinking of? >> Yes.

2:57 >> What's that? >> Evals. Yes. >> Hooks. >> Hooks. Yes. Um the three that I'm thinking of before I even get to hooks

3:04 and evils, hooks and evils are important. Tests are absolutely the class of that hooks are how you run those things at the right time. Um to me the three things that I want to optimize

3:14 on the deterministic level are making sure that I have tests, making sure that there is strict typing, making sure that there are there are splinters. Those are not total statements and they are not

3:23 enough to actually get your uh agent to write good code. as an example, if you just take from this that you have to put strict typing in place, you will still be in a bad place. Um, if you don't

3:37 discourage or ban um a way of typing that removes edge cases you don't really want to see, then the behaviors that or the states that are representable in the app are going

3:52 to lead the agent to introduce them. But the thing that I want to predominantly focus on in this talk is llinters. Lintters are our way to codify um the rules in a deterministic manner that we

4:04 want to um remove from the from the codebase. Historically like before AI when you think about llinters um the way that this worked is the

4:15 trade-off wasn't clear. Writing an actual llinter took a lot of work. um it was a maintenance cost that you had to sort of sign yourself up for and you had to consider how often the issue that you

4:27 try to lint away comes up is it even worth it and if it's if it's pre pre- aents and those are humans that are working in my codebase and the only thing I'm doing is once every free week

4:40 telling people to use the right component from the design system then I don't necessarily need a llinter for that right and people have memory so if I tell them once or twice they will

4:50 usually remember that. But then this shifts entirely into this new world where the things that are writing code do not have any memory and are going to make the same repeated mistakes over and

5:02 over again. And on the on the at the same time to write a new lter is very cheap and very easy. You can just have the agent do that and it's going to be much less frustrated with the actual

5:14 lint rules that you're trying to to write. So my claim for this is simple and if you take one thing from this entire talk that I'm going to yap about a little bit more is you can after this

5:24 talk take your laptop go to your agent and ask hey look at all my past transcripts look at my GitHub reviews the feedback from the bots the feedback from my peers and try to codify which

5:38 ones of those are actual high quality lint rules that are deterministic and will be able to capture issues that we otherwise reprompt all the time. Um, that's something that I do now and I

5:50 happen to do that even more often uh sort of on the schedule to try to see whether there are new things that I haven't captured. So, here are some examples of what I

6:00 mean when I say custom lint rules. Those are uh lint rules that try to capture conventions that you want in your codebase. There is like a whole simple class of them. Those are the ones that

6:11 you know and have written for a long time. And those are things like I want to lint out console log statements from a particular project or I want to make sure that if I have a design system or

6:23 react practices that I want to follow or discourage those are codified and are limited away. Those are the simple things but I think with agents we are getting into more interesting

6:34 complexity. We can push this a little bit further. So one example here, this is a little bit of a tricky example, but um I'll go through it. Sentry is a pretty old codebase. We are on Django,

6:46 we are on DRF. We don't have strict typing everywhere. This is a very common case for any mature codebase where you're not going to be at the top leading edge of the technology. Um but

6:56 we still want to improve the state of the world for for the agents. So concretely here um there is there was a lot of drift of what the open API schema for sentry said it's going to do and

7:09 what it did and so we started introducing lint rules for that. First lint rule is if something is a end point endpoint a function for an endpoint it needs to return a response type and

7:21 within that response type there cannot be any types so that we can infer the actual information of the type. Second, if it's a endpoint, you get that little extent schema decorator so we can ensure

7:32 that it's documented. And then third, enforce that the actual type being produced here agrees with what the open API schema at the end of the day um declares. And that helps us because now

7:46 APIs are a little bit more in sync. And that gets us to even more insane examples of um of llinters. So sir agent is a debugging agent and it has to talk to Sentry a lot. It uses Sentry

8:02 telemetry to debug and RCA issues and provide root cause analysis solutions all of that because it has to talk to open to um Sentry API we oftent times have to guide it and the way we do

8:16 it is through skills. oftentimes those are old endpoints that we are not able to um easily fix and they have some complexities that are hard to describe on the level of schema itself. So

8:26 instead we guide that through skills which is a pretty common pattern in the agentic world. But here is an example of a skill that tells how a metric should be created at sentry.

8:37 We generated a lot of those skills and examples with those code blocks as the as the examples of code that can then be run in the sandbox by agent. And in the original attempt to do that, we lost the

8:49 accuracy of the agent significantly because lots of those examples were hallucinated. So the first lintter we created was a llinter that says if there is a code block inside the skill, you

9:01 got to extract those code blocks, you have to make sure the type check, they can run in our sandbox and whatever they declare agrees with open API schema from Sentry and that helped. But what what

9:14 the f the first thing that the agent did when we tried to rerun the skill generation with that rule is oh there are lint rule errors let me remove the code blocks

9:25 and that wasn't very desired by us. So we said no no no examples have to be put in code blocks. So it put the examples back but it removed the bactics because only the bactics are linted. So let's

9:37 let's fix the problem right. So we're like, "No, no, no. If it looks like a code, smells like a code, if it has underscores and parenthesis, push those into codelocks." So I was

9:48 like, "Okay, so you want me to put code into codelocks. I'm going to describe the APIs." So now in pros, it started describing what the APIs should do, but

9:59 it did one mistake, which is it referred to the URLs uh of the APIs. So then we linted away the URLs and we played this little whack-a-ole for a little while, but at the end of the day, we ended up

10:11 with skills that are synced to our open API schema and that wasn't that expensive to do. And we have I'm not going to say unbreakable model, but we have a model where when API

10:24 changes, we actually don't have to follow up on the agent skills all the time. it it's it's kept in sync through the through the llinters. And when it comes to the actual

10:35 movements or the actual how do we do this you can use any sort of tool that you're using already es oxlin flake 8 clippy as grabb as grabb is great uh because it's language agnostic so it

10:48 kind of you know unified tooling warm fuzzy feelings about that um there is a bunch of new tools that are popping up that are trying to approach it from a different uh perspective one example is

10:58 taskless where is instead of writing the actual lter you write what you want the llinter to do. You write your intent in English language and then the actual llinter is um derived from it. Um and

11:10 then the last thing that I uh that I will repeat is try asking the agent for the actual lers. You will be surprised uh what it can what it can generate. But

11:21 that's only half the picture. What if there are rules that are deterministic cannot express um cannot be expressed through llinters? There is plenty of those, right? Those

11:33 are all types of qualitative measures like you have a state machine with just three states and I created a small PR to add small functionality and ballooned up the number of states into 10. I don't

11:46 know how to lint against that. I don't think I reliably can link against that. I can like create small tactical things like no more than four states, but it's all It's all made up. This is

11:58 a this is the situation where we're talking about qualitative measures and um this is where you are engaging your engineering experience or if tweet as as Twitter would say your taste um and you

12:11 just have to decline it straight up or you can have an LLM agent that is going to do a pass and for simpler thing help you with that concretely here

12:25 I've uh I've started to create policies for the repository and those policies are such as I don't want too many tests, I want stronger tests or I do not want a code that's too defensive. I'm sure

12:40 you've seen those where every single possible case is try code by the agent and it's it's just sloppy long long functions. I would rather have code that in its type system does not allow to

12:54 express states that are undesired. Um so I created those policies and then I created a skill that I called grumpy engineer. We call it Garfield internally and the way it works is it takes all the

13:08 policies and run sub agents for each of the policies and the sub agent can either give feedback or stand down and we loop that and this is the closest you will hear me doing loops by the way. uh

13:21 we will loop that until the agent stands stands down and says I I'm okay with that I validate that this is fine and only then I read the code and sort of my metric of success here is how much I'm

13:33 casting up the agent how much when I'm reviewing the code I'm like oh yeah this is not ideal but I can I can steer it here and I don't have to do very elementary feedback all the time it is

13:44 still qualitative but the goal here isn't to remove the code review quite yet you So this things will change as models get better but more so to raise the floor and make it easier to um to

13:57 work with the agents uh as you as you're going forward. There is also one thing that I want to mention that is um beware of a snake oil. We are talking

14:10 about the tricky part being qualitative measures and I've done experiments where I tried to capture a qualitative measure in a quantitative metric. So two concrete examples one of them is um

14:24 cyclomatic complexity and the other one is test coverage. Cyclomatic complexity uh is a measure that says how maintainable and testable the code is and it tries to base that on how many

14:36 branches happen in the code and how deep the call stack goes and um and test coverage you know test coverage tries to tell you how much of the code is covered by test. Um, and the problem with that

14:49 is when you give a number to the agent, the agent will optimize the out of that number. And it did. And I I got code bases, I got projects where I got to like a 100%

15:01 or close to 100% of test coverage. But the actual tests are bad. They are like asserting particular string in particular output of the agent. They are not measuring anything of value. Um, and

15:12 more so they are giving you a false sense of security. So be aware of that. beware of um the complexities uh of of actual qualitative metrics. Um and that's it. That's that's my call.

15:26 My call to action for you is after this try to ask your agent for llinters and policies and see whether that helps you. Um try things and make the repo remember. And that's that's my time.

15:37 Thank you.