Engineering
autoresearch avatar

autoresearch

Autonomous multi-team codebase improvement agent with specialized modes: narrow (goal-directed), broad (hypothesis-divergent), and sweep (quality-focused).

Introduction

Autoresearch is an autonomous agent system designed to improve codebases by coordinating up to three distinct teams—Red (problem discovery), Green (fix implementation), and Refactor (code simplification)—using clean-room information barriers. It operates on your codebase via a branch-based workflow to ensure safety, verifying tests at every cycle to prevent regressions.

  • Narrow Mode: Targeted optimization for specific, measurable metrics such as p99 latency, coverage percentages, or bundle sizes. It requires a baseline, a target, and a priority-ranked list of strategies to achieve the goal.

  • Broad Mode: A hypothesis-driven approach for exploring improvements without a pre-defined metric. It generates diverse hypotheses across categories (obvious, bold, creative), executes them on independent tracks, and provides a comparative analysis to determine the best path forward.

  • Sweep Mode: The default operational mode for general quality improvements, focusing on bug hunting, dead code removal, and test coverage expansion via standard iterative cycles.

  • Automated Stack Detection: Supports multiple languages and build systems including Go (go test), Node/TS (npm test), Rust (cargo test), Python (pytest), Ruby (rspec), Java (maven/gradle), Elixir (mix), and PHP (phpunit).

  • Secure Workflow: Enforces git integrity by checking for uncommitted changes, utilizing isolated branch workflows, and providing a rigorous session management system that logs cycle results to a results.tsv file.

  • Usage: Invoke via /autoresearch followed by narrow, broad, or a path for sweep mode.

  • Configuration: Customize behavior by adding a .autoresearch.yml file to your repository root to override test commands, include/exclude patterns, and team participation.

  • Resumption: Sessions are tracked in a persistent directory; use /autoresearch resume to continue interrupted work based on the last recorded state and skill version.

  • Compatibility: Compatible with major coding agents like Claude Code and supports standard git-based development environments.

Repository Stats

Stars
48
Forks
3
Open Issues
0
Language
Shell
Default Branch
main
Sync Status
Idle
Last Synced
May 3, 2026, 11:24 PM
View on GitHub