PROTEA: Offline Evaluation and Iterative Refinement for Multi-Agent LLM Workflows

In three linesPROTEA is an interface for offline debugging and refinement of multi-agent LLM workflows. It evaluates intermediate outputs with configurable rubrics, localizes bottlenecks via workflow graph visualization, and generates targeted prompt revisions. On two production-adjacent workflows, PROTEA improves document-inspection accuracy from 64.3% to 83.9% and recommendation Hit@5 from 0.30 to 0.38.

Summary generated by Claude — human-verified