@july
For agentic engineering, I found myself increasingly having to make verification gates, where I set almost certain rules that they need to be verified in order for pull requests to merge. Essentially, I need ways for the work to be checked without me actually being there. That's how I've noticed I can scale beyond just three to four PRS.
The pull requests, or rather, the number of pull requests that I'm working on, essentially can increase, but the number of agents that have my attention at a time are still constant because my RAM is almost fixed. It doesn't mean that the number of PRs that I'm working on at one time is the same. I can expand the number that I'm working on because certain pull requests will be at different stages of, let's say, merge conflicts and other things.
The other thing too is when I'm working on larger, bigger tasks, I think a lot about how do I manage pieces of work that need to be worked on? Like, I'll take a gigantic piece of task. I'll come up with a club plan, and then in that plan I'll say, okay, I want like four or five pieces that won't conflict with each other and can be merged safely. Then that will allow me to essentially make tasks composable and mergeable into main at a later time.
Because the other thing that's interesting is that, as the number of pull requests increases, the size of the surface area of the code that conflicts increases. There's no real shared memory about how things need to work, so that the space or the area where the code is changing increases. Also, the verifiability of whether that's actually a good solution or not is increasing, because the size increases.
I think there are two things:
- Enforceability needs to increase.
- You also overall want to increase the velocity of development, but you also don't want to introduce any garbage code and you want to keep the quality high.
I think pull requests, code review, and getting these things merged all become increasingly important, like DevOps is important. I've actually made my own bot, an AI review bot, that helps me merge code, because I need it to review my code and then solve all the other outstanding issues
Such as:
- Rebase main
- Fix merge conflicts
- Address bot review issues, whether they're false positives or whether they are real issues that need to be fixed
- Test them in some capacity to make sure that it works
- Where are we missing test coverage
- If everything is fixed, merge