Agent OS turned its own GitHub issues into merged PRs for 24 days
At a glance
kai-linux's Agent OS watched a project board, dispatched Claude/Codex/Gemini/DeepSeek agents in isolated worktrees, and in 24 days shipped 288 commits with 59+ merged PRs at ~8-minute median cycle time. This workflow is listed under multi-agent, pull-requests, ci, worktrees, agent-ops.
- Agents / tools
- Claude, Codex, Gemini, DeepSeek, GitHub
Building a tool for these workflows? See weekly rates.
What they did
Agent OS polls a GitHub project board every 60 seconds, hands the top ready issue to an agent in an isolated git worktree, runs tests, opens a PR on green results, and squash-merges when CI passes (or files a debug issue on red). A weekly log analyzer and backlog groomer file improvement tickets from metrics, including a documented recovery after PR #98 caused cascading CI failures.
How it went
Over 24 days on its own repo: 288 commits, 59+ PRs merged (91% merge rate), 79+ issues closed (92% closure), 122+ agent tasks, 55.7% first-attempt success, median PR cycle ~8 minutes.
Worth knowing
First-attempt success was only about 56%; the case study highlights failure recovery (structured frontmatter for failed checks) as much as the happy path.