from AgentOps311
Autonomous improvement loop that selects high-value work, runs a full research-plan-implement (/rpi) cycle, validates changes, and repeats to compound repo impr
Evolve provides a fully autonomous, operator-driven improvement loop: it selects the highest-value work (pinned queues, harvested findings, failing goals, testing gaps), runs a research/planning/implement/validate (/rpi) cycle, applies a regression gate, and repeats until a stop condition. It's designed to iteratively harden and improve codebases by automating selection, execution, and validation.
Use Evolve when you want continuous, supervised autonomous development: landing bug fixes, tightening validation, generating tests, performing audits, or processing a roadmap (.agents/evolve/roadmap). It is suitable for day-time operator runs where code changes are expected and validated each cycle.
Best fit for agent platforms that support code changes and lifecycle tooling (AgentOps CLI, RPI engine). Commonly used with Claude Code, Codex, Cursor and similar operator-capable agents.
Evolve is a goal-driven autonomous improvement loop skill for the AgentOps ecosystem. It orchestrates a full research-plan-implement cycle per iteration with fitness scoring, regression gates, and work selection ladders. The SKILL.md is exceptionally detailed with 7 execution steps, 10+ reference docs, and comprehensive flag support. The bundled validate.sh script failed entirely in the sandbox because it resolves SKILL_DIR relative to the script location but runs in an isolated temp directory where the skill files don't exist — a path-handling bug that would also affect real-world execution outside the skill's own repo.
Well-architected skill with thorough documentation and clear operational contracts. The validate.sh script has a fundamental path resolution issue — it uses dirname-based SKILL_DIR resolution which breaks when the script is copied/run from a different location. Shell injection risk in validate.sh via unquoted variable expansion in bash -c strings. The skill is tightly coupled to the AgentOps ecosystem (ao CLI, beads, /rpi) which limits standalone usability significantly.
