surpradhan

Systematically debug a failing multi-agent workflow from a captured trace. Use when an agent system produces wrong outputs, gets stuck in loops, makes poor tool choices, or fails non-deterministically. The skill refuses to propose fixes without a captured trace because debugging without observation is guessing. Works with LangGraph, CrewAI, AutoGen, OpenAI Agents SDK, and custom multi-agent setups

Set up a controlled RAG evaluation harness. Use when starting a new RAG benchmarking project, adding a new retrieval architecture to an existing benchmark, or evaluating a retrieval change against a stable baseline. The skill enforces fixing all components except the architecture under test, which is the methodological prerequisite for a meaningful comparison.