Self-Healing Codebases Are Coming… And AI Debug Agents Will Lead the Charge

something just broke in your codebase and now you need to go figure it out or you're working locally and you got some errors. We talk a lot about AI agents and building net new stuff, but I don't think we're thinking enough about how we deal with errors and logs and tracing both when we're at runtime and then when we're in prod. So, let's talk about it. If you didn't know, I'm partner and I do a daily upload. This is that channel. If you want to see more in-depth content, you can go to my main channel. But I did some exploration on this where I'm thinking to myself like, how do we make this better? How do we make a self-healing agentic system? And I think this is a lot of what folks at Meta are doing and where you're basically building the tooling for your stack. You're personalizing it. And so that way in this case we have this cohort which is net new and I think a lot of people if we were to break the attention apart I think like 90 maybe like 95% of what we'll go with 90% of the attention is on building for 0 to 1 right net new then I want to talk more like I said about errors and like ongoing slash ongoing dev cuz at some point once you get this out the door this is where it matters, right? And so in this case today what we do is we have different providers for this like you might have New Relic, you might just have a bunch of manual like logs logs you might have stack tracing it's all this observability right observability. So how do we take observability and tie it to agents so that you have something that is selfhealing? So as I was looking through this is where I landed and we had a really good comment by someone who joined our community yesterday that so the way that it would work is a user would trigger the bug. So this can either be in a prod environment or dev. Then Dino, which is what we'd be using in this case because it has open telemetry built in, which is huge. That's the sauce where open telemetry shoots everything down that you could possibly need for debugging and it's native to Dino. And if you're not familiar with Dino, Ryan Dah, he invented Node and then he also invented Dino. And he had a great talk on this if you're curious about it just about how we have the humble console log, but how having telemetry and Open Tel built in is just such a game changer. And so to continue down that path, oops, basically Vendino would emit the telemetry span and then Prometheus, which is a time series database. You'd see the things basically streaming down in real time would detect any sort of error spike which we can use with a set of rules. Oh, in the case of a scheduled YouTube uploader, it could see if there's a bunch of errors that are occurring that are abnormal and then that would fire the web the web hook to an agent API. In this case, we're going to be using Google's agent ADK and that set of agents would have all of the context needed. It would read the source of the files and the error trace. Then it would generate a new patch and then run a test suite against it. And if the tests pass, then depending on the environment, it would decide which way to go, which is really cool. And I think this is where a lot of the companies already are. But I just wanted to give you like the high level. Now some of the see like prerexs would be that you're using Dino because I think it's just better for this use case but essentially you end up with a self-healing AI assisted app and this would be like the future framework basically. So pieces of this let's see it would look something like this. This doesn't include any of the TDD which is of a test-driven development. But in our case, we'd have a front end which is client with V and Tanstack. You'd have all the different hooks and API calls. Then you'd have the back end which is Dino. So you'd have your routes. You'd have a health check as well as a web hook and then the telemetry utility and that's the magic. And then you'd have the agents which have the self-heal agent. This should probably have sub agents as well. It's a really good video on agents with the Google ADK if you're curious by a guy named Brandon. He just did a three-hour master class. You can check it out here, AI with Brandon. And I just think this is where it goes. So, it's some it's going to be some combination of the self-healing agentic workflows with test-driven development. And I just am really excited to see where this goes. You could have an optional graphana dashboard so you can see what's going on and really just keep keep on top of it. So auto patching in prod is definitely something I'm excited about. I'm not working on it actively. I just spent about I don't know hour or so yesterday thinking about it. And then in our community which is half off right now. You can check out some of the stuff that we were like basically just broke down. Let me see. It's in general. It's broke down like how the system detects bugs in real time through telemetry. Then triggers an AI agent to automatically diagnose, fix, and test the issue without human intervention. During development, it can patch code live. In production, it creates poll requests that pass CI before going live. Turning the typical debug cycle from hours into seconds. Let your software continuously improve itself. So, it would just be this constant live stream of how it's doing under the hood. logs, errors, timings, traces, status codes stored into Prometheus with a large raft, right? And so it looks something like this for the infra and then we talked about that part earlier, but yeah, I think if you tied in even like some of the MCP stuff with TDD where it's like you could have your own codebase have its own index for rag. We were talking all about that stuff last night. That's all for today. If you enjoyed the video, make sure you like the video. I'm waiting. That's your gift to the random guy on the internet today. You can also subscribe. That's your best way to support me. And if you're interested in joining the community, surrounding yourself with other people who are pushing the edge of what's possible with AI, check out Vibe with AAI. Half off right now. And I'll see you in the next slide.