Investigation agent
An agent that examines failed simulation runs and checks the evidence before it acts.
Scientific model runs fail for many reasons, and the clues are spread across logs, configuration files and outputs. At NASA’s Community Coordinated Modeling Center, I built an agent that works through that evidence.
Deterministic processing reconstructs what happened first. The agent then uses read-only tools to answer what’s still uncertain, ties each conclusion back to the source files, and returns a structured assessment. Uncertain cases go to a stronger model or a scientist for review.
Make the conclusion inspectable
A diagnosis is only useful if a scientist can trace it back to the run. The goal isn’t just an explanation. It’s one that can be checked independently.
Agentic Fault Analysis for CCMC ROR Simulations, my public presentation of this work.