Self-Evolving Meta Skill System

Problem

In multi-skill agent systems, routing rules, evaluation criteria, role definitions, and exception handling often live in prompts, config files, and tribal knowledge. When a new task pattern appears or one failure mode recurs, teams must manually inspect logs, rewrite prompts, and judge whether a change is safer or merely different. meta-skills targets that gap: it does not answer end-user requests directly. Instead, it treats the policies that control skill selection, evaluation, and optimization as versioned objects that can be revised from observed behavior.

How it works

The system turns policy iteration into a closed loop:

• Triggering: starts optimization after scheduled intervals, event signals such as a dropped success rate or new error patterns, or explicit user feedback.
• Reflection: uses an LLM to analyze recent execution logs, surface routing defects, evaluation misalignment, uncovered task types, and produce structured recommendations.
• Mutation: generates candidate policies using operators such as CLARIFY, EXEMPLIFY, ROLE_SET, THRESHOLD_ADJUST, RULE_ADD, and DECOMPOSE.
• Validation and rollout: runs candidates against a standardized test set with real LLM calls, compares outputs, latency, and token cost, then moves through shadow mode, A/B testing, and full deployment.

It also includes a policy loader, execution sandbox, append-only logging, rollback management, and permission controls to keep mutations bounded.

Boundaries

This skill fits systems that already have logs, measurable outputs, and policies that need repeated refinement. It is less suitable for brand-new skills with little history, one-off configuration fixes, or highly subjective quality judgments. Practitioners should maintain a valid test set and account for the extra LLM calls required for reflection and validation.