evo-hq/evo

What it solves

evo is an autoresearch orchestrator that automates the process of optimizing code. Instead of a developer manually tweaking a codebase to improve a specific metric (like speed or accuracy), evo allows an AI to autonomously run experiments, test them, and iteratively improve the code until it reaches a peak score.

How it works

It functions as a plugin for existing agentic frameworks (such as Claude Code, Cursor, or Codex). The system follows a loop of discovery, experimentation, and selection:

1. Discovery: The discover skill analyzes the repository to identify what to measure and sets up the necessary evaluation benchmarks and "gates" (safety checks).
2. Parallel Experimentation: The orchestrator spawns multiple subagents, each in its own isolated workspace (git worktree or remote sandbox). These agents formulate hypotheses, edit the code, and run benchmarks.
3. Shared State: All agents share failure traces and discarded hypotheses to avoid repeating mistakes.
4. Selection: Using strategies like argmax or epsilon_greedy, the orchestrator decides which successful branches to extend in the next round.
5. Gating: Every experiment must pass defined gates (e.g., regression tests) to be kept, preventing the AI from "cheating" the metric by sacrificing correctness for performance.

Who it’s for

Developers who want to automate the tedious process of performance tuning, bug fixing, or metric optimization within a large codebase using AI agents.

Highlights

• Tree Search: Uses a branching tree of experiments rather than a simple linear path to explore multiple optimization directions simultaneously.
• Multi-Backend Support: Experiments can run locally via git worktrees or in remote sandboxes like Modal, E2B, AWS, and Azure.
• Cross-cutting Scans: Employs scan subagents to find compound failure patterns across multiple traces.
• Broad Compatibility: Integrates with a wide variety of AI coding hosts including Claude Code, Cursor, Kimi, and others.