Drasko Profirovic presents Medic, an AI-assisted system for investigating failed Apache Spark jobs at Pinterest. An initial design exposed diagnostic tools through MCP to a single ReAct agent, but large logs, long prompts and accumulating instructions made the investigation noisy and difficult to maintain.
Drasko Profirovic describes tracing with OpenTelemetry and Langfuse, replayable fixtures and offline evaluations as ways to understand and compare agent behaviour. Log clustering and fingerprints help surface relevant evidence; targeted retrieval reduces red herrings, while specialist analysis can examine time-series charts within a bounded context.
Drasko Profirovic then outlines a LangGraph-based arrangement with triage, research, supervisory and healing responsibilities, informed by runbooks and previous fixes. The talk describes design choices and observed implementation problems, not a quantified guarantee that autonomous fixes will always be correct. Extending the approach to other query engines remains future work.
Watch on YouTube
