Running coding agents like engineering orgs (twitter.com)

A recent experiment by a tech director involved organizing AI coding agents, specifically Claude Code, in a structure akin to traditional engineering organizations. This arrangement included a central tech lead overseeing various teams, which comprised staff, senior, and junior agents, each assigned specific tasks. Over five weeks, the agents demonstrated remarkable autonomy, effectively collaborating, escalating issues, and completing tasks with minimal direct oversight. The results were impressive: nearly 2,900 agents participated, resulting in 1,711 merged pull requests (PRs) and about 50,000 messages exchanged among agents. This experiment is significant for the AI/ML community as it illustrates the potential for coding agents to replicate complex team dynamics traditionally seen in human engineering environments, enhancing productivity and collaboration. The structured approach not only improved the overall efficiency—evident from the jump in PRs per message—but also maintained quality, with PRs scoring similarly in reviews to those from a human-driven repo. By enabling agents to coordinate effectively and tackle tasks independently, this setup provides valuable insights into automating software development processes and could pave the way for more sophisticated AI-driven engineering workflows in the future.