Research and experiments loops

Karpathy-style autoresearch and keep-or-revert experiments, claim checking and competitor scans, where only measured gains survive.

Loops › Research and experiments

autoresearchHarnessTop pick

Karpathy's overnight research loop: one editable file, a frozen evaluator, one number. Only changes that beat the best result are kept.

Any agent · Stops when: the time or run budget is spent; you wake up to the best version

Single-metric experiments where only measured gains survive. Works for prompts, pages, build times and more.

Claude Code, Codex, Cursor and other coding agents · Stops when: the target is reached or the experiment budget is spent

Metric-driven propose, evaluate and keep loop, with all its state in plain markdown files.

Any agent · Stops when: the metric target or run budget is reached

Karpathy's autoresearch pattern packaged as an installable skill, for any metric you can measure.

Any agent that reads SKILL.md · Stops when: its run budget is spent

One change at a time, measured, kept only if it proves better. The ratchet never slips backwards.

Claude Code, Codex, Cursor and other coding agents · Stops when: no further gains, or the budget is spent

Labels every factual claim in a draft as proven, weak or unsupported, then fixes or cuts the weak ones.

Any agent with web search · Stops when: every claim left is backed by a source

Turns research into something publishable: a source map, a claim check, and a hard stop before posting.

Any agent with web search · Stops when: a checked draft is ready and waiting for your approval

A bounded scan of what competitors changed, with signal separated from noise and no public writes.

Any agent with web search · Stops when: the scan budget is spent and the report is written

A keep-or-revert optimisation loop built with no code, for things like email copy or landing pages.

Claude (Cowork) · Stops when: the target is reached