Use the canonical docs for installation and platform prerequisites
AgentSight is a Linux/eBPF tool and current installation, privilege, build, Docker, and troubleshooting details change with releases. Keep those operational instructions on eunomia.dev and the product repository rather than copying them into a marketing page. Once the released CLI is available, the useful product workflow begins with one selected command.
Pick a task you can recognize in the resulting timeline
A good first run is short enough to inspect manually and concrete enough that you know what should happen: review one commit, run one coding task, or invoke one tool. Broad background capture makes it harder to tell which process or path belongs to the agent. Start bounded, then expand scope only when the first session answers the expected questions.
Open the saved run before exporting anything
Use the saved-session report or web UI to confirm that the model/tool phases, process tree, file activity, network activity, and resource view are present for the task you recorded. If an expected signal is missing, fix the capture boundary first. Exporting an incomplete run simply turns missing visibility into a portable artifact.
Treat the local session as sensitive development telemetry
Depending on capture mode, a session can contain prompts, responses, commands, paths, headers, network targets, and other execution context. Keep the raw SQLite/session artifacts local unless you intentionally need to share them. Redact or reduce exports to the review question rather than attaching an entire database to an issue.
Move to the specialized workflow after the first successful run
Use the slow-run page when the problem is performance or cost, the review workflow when the problem is an AI-generated patch, the Agent Flamegraph guide when the question spans many sessions, and the agent-specific integration pages when runtime packaging or TLS compatibility matters. The first run is a baseline, not the whole operating model.
$ cargo install agentsight
$ sudo agentsight record -- claude
$ agentsight report serve
$ agentsight report export -o snapshot.json