Skip to content

sigilco/agentplugins-autoresearch

 
 

Repository files navigation

agentplugins-autoresearch

Let the agent run experiments, keep what works, and discard what doesn't.

A fork of pi-autoresearch and inspired by karpathy/autoresearch, ported to the agentplugins CLI.

The vision is simple: give an agent a benchmark, then let it try ideas, measure results, commit winners, revert losers, and keep going until you tell it to stop.


Quick start

# install via agentplugins
pnpm dlx agentplugins install agentplugins-autoresearch

# start a loop
/skill:autoresearch-create

The agent will ask for:

  1. Goal — what you want to optimize
  2. Metric — the number that decides keep vs discard
  3. Command — the benchmark that produces that number
  4. Scope — files the agent is allowed to change

Then it writes .auto/prompt.md and .auto/measure.sh, runs a baseline, and starts iterating.


How to use it

/skill:autoresearch-create — start a loop

Opens the autoresearch skill. Answer a few questions and the agent begins experimenting.

/autoresearch <goal> — enter autoresearch mode

Puts the agent into a tight loop: propose change → run benchmark → log result → keep or revert → repeat.

/autoresearch optimize test runtime while keeping correctness
/autoresearch reduce bundle size after pnpm build
/autoresearch improve Lighthouse performance score

/autoresearch export — open the dashboard

Launches a browser dashboard with a live chart and results table.

/skill:autoresearch-finalize — turn wins into reviewable branches

Groups kept experiments into clean, independent branches from the merge-base. Each branch can be reviewed and merged on its own.


What you get

Extension init_experiment, run_experiment, log_experiment tools + live widget + /autoresearch dashboard
Skills autoresearch-create, autoresearch-finalize, and optional autoresearch-hooks
Session files One .auto/ folder per project: prompt.md, measure.sh, log.jsonl, optional checks.sh and hooks/

Why this works

  • Benchmark-driven. The metric is the judge. No hand-waving.
  • Persistent. .auto/log.jsonl survives restarts and context resets.
  • Safe. Every experiment is committed; regressions are reverted.
  • Composable. One extension handles any domain — tests, builds, ML, Lighthouse, anything that prints a number.

Learn more

License

MIT

About

Autonomous experiment loop extension; harness-agnostic

Resources

License

Stars

1 star

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages

  • TypeScript 70.2%
  • HTML 15.5%
  • Shell 7.4%
  • JavaScript 6.9%