Autoresearch

Autoresearch is inspired by Andrej Karpathy's autoresearch approach of pairing a constraint, a mechanical metric, and autonomous iteration to produce compounding improvements overnight. A user gives it a plain-language goal; it derives a success predicate, confirms it once, and then loops: read context, define scope of which files may change, establish a baseline by running verification once, and then repeat one bounded change per iteration, keeping it if it passes and reverting automatically if it does not, with progress logged in TSV format.

A newer orchestrator classifies the stated goal and chains the underlying subcommands automatically instead of requiring manual invocation. The project generalizes the approach beyond machine learning to code, content, marketing, or any task with a number that can be measured, and rebuilt its monolithic instructions into a smaller routing file plus separate command files to cut the tokens loaded per invocation.

Hook guardrails are specific to Claude Code.