autoresearch

Autoresearch is a Claude Code skill that implements autonomous, goal-directed iteration by running a continuous Modify-Verify-Keep/Discard loop across any measurable domain, from ML model tuning to DevOps pipelines. Installed as a set of slash commands, it exposes 13 distinct commands including `/autoresearch:plan`, `/autoresearch:loop`, `/autoresearch:debug`, `/autoresearch:security`, and `/autoresearch:evals`, each backed by a self-contained command file of 94 to 120 lines. The skill integrates directly with Claude Code and also supports OpenCode and OpenAI Codex. On each iteration, Claude reads context, makes one focused change, commits via git, runs mechanical verification, and automatically reverts the commit if the metric regresses, with all results logged in TSV format. A notable architectural detail is the v2.1.0 rebuild that replaced an 813-line monolithic SKILL.md with a 41-line router, cutting token consumption by roughly 95 percent per invocation. Developers, ML engineers, and technical teams who want unattended overnight improvement cycles against a defined numeric metric are the primary audience.

Lo que la gente pregunta sobre autoresearch