GitHub suspends developer after AI agents fire 50 PRs in a day

The developer had been running a loop in which AI agents fixed bugs, added tests, and cleaned up documentation across 15 public repositories, then had another AI grade the results until scores improved. Subagents were assigned one per repository so all 15 ran at once. In a single day this produced about 50 pull requests, each pushed to repeatedly, with CI results checked through the API every few dozen seconds and failed runs restarted.

The suspension arrived mid-way through a fifth round of fixes, when every GitHub API call began returning HTTP 403. Sign-in was also blocked for a while, and the developer took about half a day to find the appeal form and submit a request. The appeal process itself asked only for an honest account of how the account was being used; a support agent replied several hours later asking what the developer planned to do with GitHub, and the restriction was lifted five days after that reply. During the freeze, repositories and pull requests became invisible, published sites went dark, and scheduled tasks failed.

The recovery rules the developer settled on reflect GitHub's own REST API guidance: send requests sequentially rather than in parallel, leave at least a second between write calls, and prefer notifications over repeatedly asking the API for results. Beyond those, the developer set daily ceilings, changed push behavior so a single push follows passing local tests, and instructed agents to stop and report rather than retry when a 403, 429, or secondary rate limit appears.