labrat

labrat runs Claude Code or Codex as peer operator interfaces on a defined research problem with a baseline and a consistent evaluator, using population search rather than a single thread: families of ideas compete for compute budget, and the families that produce real signal earn more room to keep exploring. An async supervisor keeps the queue moving without a global cycle barrier, so worker agents keep evaluating descendants as slots free up, while probe, mutation, crossover, and audit workers execute bounded tasks under the supervisor.

Credits are minted by stable, reproducible progress and spent on new descendants, and a family can earn extra status by winning a held-out challenge that was not already built into the local search metric, rather than only by climbing the local hill. State is carried forward through durable, append-only files rather than a database, and a dashboard shows both the current champion and the current decisive-challenge leaders.

Good problems for labrat have a clear baseline, a bounded experiment runner, a consistently scoreable metric, and at least one harder held-out challenge.