Build a self-improving skill loop: schedule a memory review that rewrites its own skills
BuildMVPFast·medium signal
Anthropic's AutoResearch-style pattern pairs a scheduled memory-review pass with automated quality improvement, so an agent periodically reads its own accumulated memory/outcomes and refines its skills as a managed routine rather than a one-off prompt. It's the production shape of 'systems that learn run-over-run' — the review cadence is the engine, not a nice-to-have. Add a weekly Routine that summarizes what each skill produced, scores it, and edits the SKILL.md to drop low-yield behavior and reinforce what worked.