Autoresearch — Self-Validating Hypothesis Loop

Autoresearch is a pattern, not a package. The shape is: agent generates a hypothesis → agent runs a bounded experiment → the experiment produces a number → git keeps the change if the number improved (and regressions still pass), reverts otherwise. Repeat for hours.

Everything runs inside Claude Code. The harness is one of two skills you already have:

• /loop — Claude reruns the same prompt on a cadence (or self-paced), driving experiments one at a time. Progress streams into the session as it happens, so you can watch the metric move.
• /agent-teams — spin up a proposer agent and an independent validator/judge agent. The proposer changes code; the validator runs the metric script + regression tests and decides commit vs. revert. Two perspectives, harder to game.