Run a self-hosted code forge where agents claim tasks and merge their own code

Joel Cone, head of engineering at a fintech startup, built Bloomright, a self-hosted git forge in Rust where AI agents are the main users, and from 8 October 2026 had an agent fleet develop it on itself. He writes Tasks with goals, acceptance criteria and expected paths. A lead Claude Opus agent plans work, writes briefs, picks which model does what and watches CI. Up to three worker agents claim Tasks through time-limited leases with narrow tokens, each on its own clone with a throwaway Postgres; GPT-family models take well-specified features, Opus takes critical-path code. Agents push changes with plain git, CI runs on home lab machines, a separate AI reviewer returns a typed verdict, and a merge queue lands one squashed commit per Task.

By 2 November the fleet had put 298 commits on main from 430 proposed changes and written 24 design records. A second product on the forge landed 116 changes in 14 days, 111 by agents, with a 92.9% pass rate over 1,905 CI jobs. From 20 October, green CI plus an approving review merges a change unless it touches a protected file; Cone still approves every production deploy. An independent second review caught a database race that tests and the first reviewer missed. He stayed a bottleneck: one night the fleet sat idle about 8 hours waiting on routine approvals.

The hard limit was a cheap DRAM-less SSD, not tokens. The disk filled and stopped three agents at once, and five lanes stalled for an hour on I/O pressure while dashboards looked healthy. Fixes were frequent trims, I/O priorities per lane and never running CI on agent machines. The post launches a waitlist and the harness behind the models is not named.