How can Agents turn execution results into reusable experience? Drawing on official implementations from Claude Code, Codex, and OpenClaw, I explore an experience-building system that supports evaluation and rollback, from reflection and validation to persistent memory, dynamic retrieval, and background curation.
An Agent's plan is more than a to-do list. I explore how Plan-and-Execute, feedback and replanning, multi-path search, SOPs, and state machines work, along with their engineering trade-offs, execution state, idempotency, permissions, and evaluation.
Wood Chen