Slopless in Atlanta: Six Steps to Agent Pull Requests That Maintainers Actually Merge

Any codebase you depend on is now yours to fix, even in a language you don’t write. The problem is that maintainers are drowning in sloppy pull requests blasted at them from other people’s clankers. This talk is a method for contributing to any software project using coding agents without producing slop, shown on one rather nasty real bug. You can run the same steps on the next bug in your own codebase.

Handy is the open-source dictation app I use to talk to my agents, with 31,000 stars on GitHub. Its memory grew from 249 MB to 9.5 GB in 36 hours and the kernel killed my machine mid-sentence. The backend is 40,000 lines of Rust, a language I don’t write. Someone else’s agent-written fix didn’t compile. Mine was merged. The difference was the engineering methodology I led the agents with:

1. Give the agents a way to run the app. A harness that runs it without me, including a way for an agent to speak into the microphone.
2. Record every layer at once. No one tool could see across the process boundaries the leak crossed, so the agents built one.
3. Measure, don’t guess. Eight possible causes. The measurements cleared seven.
4. Treat each fix as an experiment. Three candidates, each behind a switch. One did nothing and was deleted.
5. Have agents who know nothing review it cold. They showed the headline metric hid most of the leak. The fix was a big reduction, not a cure.
6. Ship the proof with the patch. -2 MB over four hours against +3,885 MB without it.

None of this is specific to Rust, memory leaks, or open source. I’ll show the whole arc, from the machine dying to the merge, and explain what I’d do differently next time. The full write-up is at gallon.me. Point your agent at it and it will know how to fix a bug this way.