autoresearch — Claude Code plugin: experiments and setup
This plugin runs unattended experiment loops: it edits one file, scores the result, keeps gains, and reverts failures. Use it when code has a measurable outcome.
/plugin marketplace add anthropics/claude-plugins-community
/plugin install autoresearch@claude-communityWhat autoresearch does
autoresearch scans a project for tunable code, scoring logic, magic numbers, and prompt templates. The /autoresearch:autoresearch-discover skill ranks possible targets and suggests metrics. Once you choose a file, /autoresearch:autoresearch path/to/file.py reads it and asks how the result should be measured.
The main skill prepares an experiment harness with instructions.md, an evaluation script, a test-data template, and a launch prompt. The agent is limited to the selected file, keeps changes that improve the score, and rolls back failed attempts. The plugin also includes templates for the instruction, evaluation, and launch files, plus references/lessons.md.
A numeric metric is required, and noisy or non-deterministic evaluation can undermine the loop. The plugin targets Claude Code and works around a file that can be safely isolated for edits.
What it helps with
- scan a codebase for tunable candidates
- choose a numeric score for an experiment
- generate an evaluation harness for one file
- keep better variants and revert failed changes
- run repeated trials without manual intervention
Example prompts for the agent
- Run
/autoresearch:autoresearch-discoverto find code that can be optimized. - Set up experiments for
path/to/file.pyusing a numeric metric. - Create the evaluation harness and launch an unattended loop for this file.
How to install autoresearch
Plugin marketplace
/plugin marketplace add anthropics/claude-plugins-community/plugin install autoresearch@claude-communityStep by step, with commands for each platform: Claude Code plugins: marketplaces, installing, and building your own
Questions about autoresearch
How do I install autoresearch in Claude Code?
Run /plugin marketplace add anthropics/claude-plugins-community, then run /plugin install autoresearch@claude-community.
What does autoresearch do?
It finds tunable code, builds an evaluation harness, and hands repeated trials to an autonomous agent. Better changes stay; failed ones are reverted.
What kind of metric does autoresearch need?
It needs a measurable numeric metric and a stable evaluation process. Network calls, randomness, and variable timing can make the loop chase noise.
