Auto-Evolving Agent

Problem Being Solved

Agent capabilities often lack a stable self-improvement loop: SKILL.md, triggers, SOUL.md, workflows, and prompts still rely on manual iteration. This skill frames agent capability maintenance as a controlled evolution process, helping developers explore alternatives, compare options, optimize from feedback, and reuse validated patterns for OpenClaw.

How It Works

• Hybrid evolution strategy: combines Darwinian exploration for generating candidate solutions and Lamarckian optimization for refining existing paths from feedback.
• Five-ring architecture and triggers: supports user-initiated evolution, automatic reflection after tasks, and optional periodic evolution, covering SKILL.md tuning, new skill creation, workflow optimization, and SOUL.md refinement.
• Multi-objective scoring: evaluates proposals across feasibility, expected impact, risk level, implementation cost, novelty, and rollback safety, with tiered confirmation for higher-risk actions.
• Gene pool and elite patterns: stores successful high-fitness patterns in a Gene Pool, promotes repeatedly validated and broadly applicable patterns into an Elite Pattern Pool, and crystallizes qualifying patterns into standalone skills.
• Safety mechanisms: emphasizes transparency, controllability, reversibility, incremental change, and user-in-the-loop decisions, with Git snapshots before important changes and line-by-line confirmation for high-risk operations.

Boundaries

Best for developers who want structured capability iteration and are willing to review, confirm, and roll back changes. Not suitable for unattended system modification, file deletion, dangerous commands, or bypassing audit controls.