Public recorded session

Inspect a causal run profile without turning the demo into a benchmark claim.

The public demo shows how AgentSight organizes a saved session. This page explains what to inspect, how the layers relate, and what the interface alone does not prove.

Verified with AgentSight v0.2.67Reviewed August 5, 2026Maintained by Yusheng Zheng

AgentSight live sessionspublic demo

saved sessionlocal analysismultiple evidence views

Open the interactive demo Inspect the source

Intent and turns

Identify the selected task, turn boundaries, and the model or tool decisions that frame later system activity.

Model activity

Review call timing, token-related aggregates, and provider traffic as one layer rather than the entire explanation.

Process family

Locate commands, child processes, status, waits, CPU, and memory phases associated with the recorded run.

System effects

Inspect file activity and network destinations, then connect each effect to the task step that made it relevant.

Question

What can the saved session support?

It can support statements about the events and aggregates present in the selected record. It can help a reviewer reconstruct likely causal paths and choose a follow-up experiment. It does not automatically prove that an unobserved event never occurred, that a destination was malicious, or that one run represents every version and environment.

Review sequence

Move from the task to the effect, then back to the decision.

Begin with the run boundary and intended task. Identify an expensive, risky, or unexplained phase. Follow the process and effect evidence supporting that phase, then state the smallest conclusion and the next verification step.

Terminal

$ sudo agentsight record -- <agent-command>
$ agentsight report audit --json
$ agentsight report export -o review.json