Braintrust and Basis open-sourced Trajectory Behavior Specs, a Markdown-based standard designed to evaluate hundreds of intermediate decision steps in long-horizon agents against production traces.
Key Takeaways
- ✓Moves beyond binary end-result evaluation to assess step-by-step decision quality across long trajectories;
- ✓Markdown-based specification language bridges product requirements and automated LLM judge prompts;
- ✓Continuously tests production traces against specs to generate automated regression suites during agent refactoring.