When LLM agents treat your codebase like a science lab

What it does

Evo plugs into agentic frameworks like Claude Code or Codex and runs an optimization loop on your codebase. It first explores the repository to figure out what to measure, then spawns parallel subagents in isolated git worktrees to test edits against benchmarks. Experiments that improve the score are kept; failures and their traces are shared across agents so the next round does not repeat them.

The interesting bit

Instead of a single greedy hill-climb, Evo maintains a tree of experiments and picks which branch to extend using frontier strategies like Pareto sampling or epsilon-greedy exploration. Between rounds, scan subagents read trace batches in parallel to surface compound failure patterns, giving the swarm a shared memory of what not to try.

Key highlights

• Tree-structured search with multiple frontier strategies (argmax, top-k, softmax, Pareto-per-task)
• Parallel subagents in isolated git worktrees, each reading shared failure traces and annotations
• Gating system that discards experiments passing benchmarks but failing correctness or safety checks
• Automatic benchmark discovery and instrumentation via the discover skill
• Runs locally or on remote sandboxes (Modal, E2B, Daytona, AWS, Azure, SSH)

Caveats

• Codex integration requires manual hook trust or a special install flag, adding friction compared to other hosts
• The CLI, skill files, and hook protocol must stay in strict version lockstep; drifting them caused silent failures in earlier releases

Verdict

Worth a look if you are already using Claude Code, Codex, or Cursor and want systematic, measurable optimization without hand-holding every experiment. Skip it if you are not using a supported agent host, or if you prefer deterministic refactors over autonomous trial-and-error.