Agent Observability Needs a Profiler, Not Another Tracer

arXiv 2609.20301·medium signalXBlueskyLinkedInCopy link

arXiv 2609.20301 (submitted 14 Sep 2026) argues existing agent observability tools handle per-execution debugging and tracing but cannot answer cross-run questions: where failures concentrate, what triggers unsafe effects, which tasks consume the most budget over days and weeks. AgentPProf borrows the systems-profiling model of attributing resource consumption to responsible code paths, except the responsible entity is task intent such as diagnose authentication or compare branches rather than a stack frame. For anyone running a standing fleet of agents, the distinction between debugging one run and profiling a thousand is the operational gap this names.