The Flight Logs
Every log is one agent, one task, one outcome, written by the rider who held the reins.
THE CORPUS
932,791 pull requests opened by AI coding agents — OpenAI Codex (OpenAI), GitHub Copilot (GitHub), Cursor (Anysphere), Devin (Cognition), Claude Code (Anthropic) — across public GitHub repositories, classified by outcome with merge and burn rates for each agent. Browse by agent, or by repository.
| Agent | Maker | Flights | Merge rate | Burn rate | Repos |
|---|---|---|---|---|---|
| OpenAI Codex flights | OpenAI | 814,522 | 94% | 0% | 84,702 |
| GitHub Copilot flights | GitHub | 50,447 | 74% | 2% | 14,492 |
| Cursor flights | Anysphere | 32,941 | 90% | 0% | 12,699 |
| Devin flights | Cognition | 29,744 | 68% | 3% | 4,747 |
| Claude Code flights | Anthropic | 5,137 | 83% | 1% | 1,915 |
Machine-readable: all agents (JSON) · all repositories (JSON) · all authors (JSON)
RIDER-WRITTEN LOGS
Hand-written logs from riders who held the reins — the task, the numbers, the burns, and the lesson.
Lesson: Read the code that assigns ownership, not just the code that checks it. The auth gates were solid; the hole was a background reconcile job handing published logs to whoever claimed a matching handle. And a confirm: true the model sets is advice, not consent: human approval has to come from outside the model.
Lesson: "Tests pass" is not "safe to ship": agent-written code needs an adversarial review pass before merge — ours caught an IDOR and a stored XSS that 38 green tests missed.
THE NEXT HORIZON
Start small. Hold the reins. Leave a record.