uditgoenka/autoresearch

What it solves

Autoresearch is an autonomous improvement engine that turns AI coding agents (like Claude Code, OpenCode, or OpenAI Codex) into a relentless loop of experimentation and optimization. It solves the problem of manual, trial-and-error iteration by automating the process of making a change, verifying it against a mechanical metric, and either keeping or reverting the change based on theoutcomes.

How it works

The system operates on a loop: it reviews the current state and history, makes one focused change, commits the change to git, runs a mechanical verification (such as a test or benchmark), and then either keeps the improvement or reverts it if the result is worse.

Key operational rules include:

• Atomic Changes: Only one change is made per iteration to ensure clarity on what caused a result.
• Mechanical Verification: It relies on numbers and metrics rather than subjective judgment.
• Git as Memory: Every experiment is committed to git, allowing the agent to read logs and diffs to inform the next iteration.
• Automatic Rollback: Failed changes are reverted instantly.

Who it’s for

Developers and engineers who want to autonomously improve test coverage, reduce bundle size, hunt and fix bugs, or perform security audits without manual intervention.

Highlights

• Multi-Agent Support: Compatible with Claude Code, OpenCode, and OpenAI Codex.
• Autonomous Orchestrator: Can take a plain-language goal and automatically derive a success predicate and pipeline.
• Specialized Commands: Includes dedicated loops for debugging (/autoresearch:debug), fixing errors (/autoresearch:fix), security auditing (/autoresearch:security), and shipping workflows (/autoresearch:ship).
• Safety Guardrails: Includes hooks to block dangerous commands, protect private keys, and prevent context bloat.
• Token Efficiency: Uses a routed architecture to reduce token consumption by up to 95% compared to monolithic prompts.