What changed
Microsoft Research introduced AgentRx, a framework for collecting trajectories, identifying failure signatures, proposing root causes and testing repairs across agent systems.
Why it matters
Non-deterministic agent failures need evidence across prompts, tools, state and environment; ad hoc log reading does not scale to recurring production defects.
What to do
Feed three failed and three successful traces into the framework, verify each proposed cause manually and require a regression test before accepting a repair.
Caveat
AgentRx is a research framework, diagnosis quality depends on trace completeness and model-generated explanations can still mistake correlation for cause.