Agent Systems: Learning and Adaptation Pattern

Intent

Use evaluated experience to improve future behavior without changing the task interface.

Outline

1. Recurring problem: repeated tasks reveal systematic errors or opportunities for improvement.
2. Structure: task traces → evaluation → policy or prompt update → future tasks.
3. Stable interface: task input and output contract.
4. Variable implementation: prompt, policy, memory, routing rule, or model.
5. Strength: the system improves from repeated use.
6. Risks: metric gaming, feedback loops, and propagation of bad memories.
7. Guardrail: never optimize a metric without checking what the metric omits.

Suggested figure

10-learning-adaptation.svg in the Agent design pattern illustration folder.