A new tool treats AI agent runs the way engineers treat slow code: something you can flame-graph.
AgentPProf is a profiler built for AI agents that now run for days or weeks, juggling tools and system resources on a single task. Instead of tracking code paths like a traditional profiler, it models agent trajectories as a "semantic operation stack," recursively splitting a run at task boundaries to attribute cost and failures to specific subtasks like diagnosing an error or comparing branches. It then aggregates those trajectories into pprof-compatible profiles, so developers get the same flame-graph visualizations systems programmers already use for CPU and memory bottlenecks. On CodeTraceBench, it hit a 0.764 B3 F1 score against human annotations, and on three problem-localization benchmarks its profiles improved MAP by up to 56 percent. The code is open-source on GitHub under eunomia-bpf/agentsight.
This matters because agent observability has mostly meant per-run debugging and tracing, useful for figuring out why one execution broke but useless for spotting patterns across hundreds of runs. Teams building long-horizon agents are essentially flying blind on which recurring subtasks quietly burn most of their token budget or trigger unsafe actions. A working cross-run profiler is the same shift performance engineering went through decades ago, moving from print-statement debugging to aggregate hotspot analysis.
It is worth remembering this is a preprint, not a production-hardened tool, and "task intent" is a far fuzzier unit to profile reliably than a function call. Whether AgentPProf holds up outside its own benchmarks is the real test.