Evo turns a codebase into an automated research loop by discovering optimization metrics, instrumenting benchmarks, and running parallel subagents that try code changes, keep improvements, and discard regressions. Each subagent operates in its own git worktree, reads failure traces and shared annotations, and can fork multiple search directions from any committed node. The framework implements greedy hill climbing with tree search rather than a single branch, enabling parallel exploration across independent agent branches.