Agents that improve their own playbooks

Can agents inspect their own failures, revise their prompts and tool strategies, and beat their previous versions on tasks they have not seen?

Our agents fail in recognizable ways: a research step skipped, a tool called in the wrong order, a brief that misses the point. Today a person reads the failure and edits the playbook.

Reflective optimization methods such as GEPA show that language feedback on failures can drive measurable improvements to prompts and workflows, often with far fewer trials than reinforcement learning.

1. 1Collect failed and successful runs with structured critiques from a separate evaluator.
2. 2Let an optimizer propose revised prompts and tool strategies, keep a population of candidates, and select on held-out tasks.
3. 3Keep the distinction explicit: these are changes to prompts and workflows, not to model weights.

Revised playbooks that outperform the versions they replaced on unseen tasks, reviewed and approved by a person before deployment.

ProposedPlaybooks are revised by hand today.

• GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning (arXiv:2507.19457)

Cited as methodological references. Playhouse is not affiliated with these authors.