self-learning-agent-patterns
Installation
SKILL.md
Self-Learning Agent Patterns
Overview
Reactive agents respond to each task independently, carrying no memory of what worked or failed before. Adaptive agents extract structured knowledge from their own experience and apply it to future tasks. Self-learning patterns define the pipeline by which raw observations become instincts, instincts become rules, and rules become skills — moving an agent from static to compound improvement over time.
When to use: Designing agents that operate repeatedly on similar tasks and should improve with use; building multi-agent pipelines with feedback loops; reviewing agent memory architectures for learning capability; any system where observed success and failure should influence future agent behavior.
Quick Reference
| Pattern | Signal | Extraction Method | Graduation Criteria | Application |
|---|---|---|---|---|
| Repeated Correction Capture | User corrects same error 2+ times | Log correction event → diff original vs corrected | 3+ occurrences, confidence > 0.7 | Promote to instinct |
| Success Pattern Mining | Task completes without correction | Extract strategy fingerprint from successful run | 5+ successful uses across sessions | Promote to instinct |
| Failure Taxonomy | Task fails or is retried | Classify error type, extract trigger conditions | 2+ failures with same root cause | Promote to negative rule |
| Preference Accumulation | User accepts output unchanged | Log format/style choices accepted without edit | Consistent across 10+ sessions | Promote to style rule |
| Temporal Outcome Tracking | Downstream metric improves/degrades | Attribute metric change to prior agent decision | Statistically significant correlation | Adjust strategy weight |