Agent frameworks & orchestration
Reliable Memory from Unstable Growth: Curating Self-Written Agent Playbooks with Randomized Experiments
Randomized Block Curation tests whether groups of self-written lessons improve task success, then selects among playbooks using fresh practice episodes.
Why it matters: For systems that reuse written lessons, the abstract suggests testing lessons’ effects on task success rather than trusting the agent’s judgment. It also makes fresh practice episodes part of choosing which playbook to serve.