autoresearch: Run Goal-Driven Improvement Loops with Coding Agents

Overview

Autoresearch is a set of agent skills and commands for running iterative improvement workflows in Claude Code, OpenCode, and OpenAI Codex. It addresses the gap between asking a coding agent for a one-off change and having it make repeated, measurable attempts while preserving useful results.

The core loop is most useful when a goal has a clear scope and a command or other mechanical check that can judge progress. The repository also includes workflows for debugging, fixing errors, security review, documentation, and shipping, so teams can use it beyond metric optimization.

Key Features

• Runs a change, verification, and keep-or-revert loop, with iteration results recorded in TSV files.
• Supports goal setup that defines scope, metric, direction, and verification before work begins.
• Provides commands for debugging, fixing errors, security audits, scenario generation, documentation, and release preparation.
• Offers goal orchestration that selects a workflow from a plain-language request.
• Includes results analysis and a regression gate for reviewing progress and checking candidate changes.
• Installs for Claude Code, OpenCode, or Codex, with shared core skill support across the three tools.
• Provides Claude Code hooks as guardrails for actions such as reading sensitive files and running dangerous commands.

Use Cases

• A developer wants to improve test coverage or reduce bundle size and has a repeatable command that reports the metric.
• A maintainer needs an agent to investigate several suspected bugs or work through existing test, type, or lint failures in bounded steps.
• A team wants to assess a change against a baseline before pushing, using regression checks to identify newly introduced failures.
• A product or engineering team needs help exploring edge cases, interrogating requirements, or iterating on a decision without a simple numeric objective.
• A project owner wants to generate or refresh codebase documentation through an agent workflow.

Project Facts

• Language: Shell
• License: MIT
• Stars: 6.5k
• Forks: 483
• Topics: ai, autonomous-agent, autoresearch, claude, claude-code, iteration, karpathy, productivity, skill
• Archived: No

Getting Started

For Claude Code, install with:

npx skills add uditgoenka/autoresearch

Restart Claude Code after installation. For OpenCode, Codex, plugin installation, and setup details, see the README.

Considerations

The loop depends on a well-defined scope and a reliable mechanical verification step. Subjective goals may need a separate workflow rather than the core metric-based loop. Iteration limits are bounded by default, while unlimited runs require explicit opt-in. The README describes hooks as guardrails, not a security sandbox, and notes that hooks are Claude Code-only; the core skill, runtime, installation, and verification surface also support OpenCode and Codex.