Start from a real run, not a synthetic dashboard. AgentSight gives you a system-level profile and skillset you can use for debugging, review, tool evaluation, and incident investigation.
The agent ran for 25 minutes and burned through tokens, but the transcript does not explain where the time went.
Break the run down by model calls, shell time, repeated scans, long-running subprocesses, network waits, CPU, and memory.
A diff tells you what changed, not how the agent got there or whether it actually ran the checks it claimed.
Attach a run profile to an AI-generated PR with commands, tests, failures, touched files, network calls, and run cost.
Claude Code, Codex, Gemini CLI, and other agents expose different logs and hooks, and some behavior happens in child processes.
Trace the agent process tree from outside the application without SDKs, proxies, or vendor-specific integrations.
Raw traces are too detailed for reviewers, managers, or tool authors. They need findings, evidence, and next actions in one page.
Use an AgentSight skill to drive the profiling workflow and guide the agent from raw evidence to a self-contained HTML artifact.
A tool says it is read-only, but the interesting question is what it does during a real agent run.
Run tools under system-level tracing and inspect file writes, child processes, external calls, and token impact.
A repo or environment is broken after a long session, and the final transcript is too coarse to identify the turning point.
Use the timeline to inspect the prompt, command, process, file operation, and network event around the failure.
Agent tools can run package managers, scripts, browsers, and cloud CLIs. Not every connection is visible in an LLM trace.
Attribute network endpoints back to the agent process family and inspect them alongside the prompt and command context.
The live demo uses a recorded Claude Code session. The reference docs stay on Eunomia for install, runtime, Docker, and OpenTelemetry details.