Arav Misra and I spent the last few months building autonomous coding agents as a side project at Headway. The big takeaway? They don't just need prompts, they need strict structure. Our first version pulled from free-text Slack messages and failed hard. Agents duplicated work and misread requests. Everything changed when we moved them into Linear. Using structured tickets with clear fields and ownership labels gave them a reliable "brain." Now, they are completely autonomous. On a set cadence, they read Linear, pick up tasks, and write results back. They even babysit their own PRs, fixing CI errors and responding to reviews. They’ve already merged over 250 PRs, and that number is rising quickly. Zero human-written code, but 100% human-reviewed. This only happened because the Eddy team built the foundation, and Headway and my manager, Robert Chen, gave me the rare space to chase a side project. It is a big reason why I love building here. We wrote this up in a Medium article here: https://lnkd.in/erB8S-PQ If work like this interests you, don't hesitate to reach out!
Fantastic insight, Will Anderson and Arav Misra. The pivot from free-text to structured tickets in Linear mirrors my own experience in RCM and operations—structure is indeed the ultimate force multiplier for efficiency. Impressive work on the multi-agent system foundation. Looking forward to following more of your work at Headway
Will - Just curious. Have you ever tried to contact your Billing department at Headway? Do your autonomous coding agents know how to actual process a problem in a timely manner through your email channels? 6 hour response time? Check.
The "strict structure over prompts" takeaway matches what we found building an AI creative studio on the marketing side: the agents that compound are the ones with tight briefs, clear evaluation gates, and a human judgment layer deciding what ships. Impressive that you got to 250 PRs in eight weeks. Curious what your evaluation gate looked like?