self-learning_skill

Overview

This skill enables agents to learn from their execution history by capturing structured traces, extracting actionable insights, and adapting agent behavior over time. It converts raw execution patterns into reusable knowledge that improves coordination, reduces recurring failures, and speeds up workflows. The capability is designed for orchestrator/worker/reviewer setups and supports session-local, agent-specific, and system-wide adaptations.

How this skill works

The skill collects structured execution traces after tasks complete, including approach, tools invoked, results, failures, and timing. It analyzes multiple traces to detect success, failure, efficiency, coordination, and knowledge-gap patterns, then generates prioritized, actionable insights. Adaptation applies insights immediately in-session, persists recommended changes to agent instructions after confirmation, or proposes system-wide updates with manual approval. A persistent knowledge base stores traces, insights, adaptations, and metrics for continuous improvement.

When to use it

• After completing a complex or multi-step task to capture learnings
• When recurring failures or recurring review comments appear across tasks
• If an agent discovers a more effective approach that should be preserved
• During multi-agent workflows to consolidate coordination patterns
• When performance metrics show degradation or opportunities for optimization

Best practices

• Capture traces automatically where possible and require minimal manual fields to lower friction
• Start with session-local adaptations and promote changes only after multiple confirming instances
• Prioritize insights by frequency and impact before applying persistent or system-wide changes
• Preserve original behavior in comments when updating agent instructions to enable rollback
• Include minimal evidence (counts, examples) with each insight so teams can validate recommendations

Example use cases

• Orchestrator: collect post-merge coordination traces and reduce worker concurrency when merge conflicts recur
• Worker: run validation hooks before completion, save trace of tools and failures for later analysis
• Reviewer: log recurring review findings to extract failure patterns and suggest workflow changes
• Task manager: analyze decomposition traces to improve future task breakdown and assignment
• Cross-agent: build a knowledge base that mandates loading core context files to prevent common errors