Skip to main content
Fuad Ali demonstrates a self-improvement loop for AI agents using a toy-shopping agent with a planted bug. The workflow traces agent behavior, uses Signal to identify recurring issues, turns failures into evaluations and regression datasets, and runs experiments to check fixes. The workshop also explores where human review has the most impact.

​
What you’ll learn

  • Trace agent behavior and use Signal to identify issues over time.
  • Turn failures into evaluations and regression datasets.
  • Run experiments to check that fixes improve agent behavior.
  • Keep people involved in evaluation design, dataset curation, and fix review.
Watch the workshop on YouTube.