Profile slow or expensive AI agent runs

A long wall-clock time or a large token bill does not tell you which decisions, retries, scans, or subprocesses caused it. AgentSight builds a run profile from the model and system layers together.

Separate model, shell, and wait-heavy phases.

Locate repeated commands, file scans, and failed retries.

Export evidence that another engineer can inspect.

Start from the expensive phase, not the dashboard

Record the real command, then inspect the process family and saved session. The goal is a short list of causal paths that explain the delay, not another collection of unrelated counters.

Use the profile to choose the next experiment

A model-heavy loop suggests prompt or tool-policy work. Repeated repository scans suggest a retrieval or context problem. Long child-process phases point to the build, test, or network path that should be measured next.

Terminal

$ sudo agentsight record -- claude
$ agentsight report export -o snapshot.json

Primary sources

AgentSight README and CLI examples