AgentPProf brings pprof flame graphs to agent trajectories by segmenting on task boundaries instead of call stacks

arXiv 2609.20301·low signalXBlueskyLinkedInCopy link

A Sep 14 arXiv paper argues agent observability tools do per-execution debugging and tracing but not cross-run profiling, so developers cannot answer where failures cluster or which tasks consume the budget at scale. The obstacle is that the responsible entity is a task intent like 'diagnose authentication' rather than a code path with a stable identifier, so the authors define a semantic operation stack and recursively split trajectories at task boundaries, then emit pprof-compatible profiles you can view as flame graphs. Segmentation reaches 0.764 B-cubed F1 against human annotations on CodeTraceBench and the resulting profiles raise MAP by up to 56% on three problem-localization benchmarks.