Arav Misra’s Post

Eight weeks ago we built a system of AI agents at Headway to clear out our engineering backlog and proactively find and fix issues. Since then it's shipped over 250 pull requests. None of the code was written by a human (although it was all reviewed by one). Our backlog had over a thousand items, some older than three years, and every one of them was a small papercut but never mattered enough to beat the next feature off the roadmap. Slow pages. A broken link nobody clicked. Tests that failed for reasons unrelated to the code. Essentially the broccoli of engineering. So Will Anderson and I built something to eat the broccoli for us. Agents autonomously find bugs, write the fix, open a PR, then babysit it through CI and code review until it's ready. What's left for the engineer is the part that actually needs a person: looking at the diff and deciding whether it ships. I went in assuming the hard problem was getting an AI to write decent code, but it wasn't. The hard problem was orchestrating dozens of agents to work in parallel without colliding, without quietly hallucinating, and without drowning reviewers in PRs nobody asked for. Turns out building autonomous swarms is guided by the same principles that make human teams work: keep the scope defined, make ownership obvious, give everyone a structured place to coordinate. My colleague Will and I wrote up how it's built, where it broke, and what running it in production for two months actually taught us. Read the article here: https://lnkd.in/e4ZuU6eR If you’re interested in building technical healthcare infrastructure and creating the future of mental healthcare in America, we’re hiring!