Systematic debugging for AI agents: Introducing the AgentRx framework
Shraddha Barke, Arnav Goyal, Alind Khare, Chetan Bansal
At a glance
- Problem: Debugging AI agent failures is hard because trajectories are long, stochastic, and often multi-agent, so the true root cause gets buried.
- Solution: AgentRx (opens in new tab) pinpoints the first unrecoverable (“critical failure”) step by synthesizing guarded, executable constraints from tool schemas and domain policies, then logging evidence-backed violations step-by-step.
- Benchmark + taxonomy: We release AgentRx Benchmark (opens in new tab) with 115 manually...