Skip to main content
A trajectory is one agent run after Bento has assembled and analyzed it. Take a raw run/trace, let Bento reason over its steps, and the result is a trajectory: the run plus Bento’s read on whether it looks suspicious, whether it’s a win, whether tools errored, and the findings extracted from it. Trajectories are the Monitoring page in the dashboard, the list of analyzed runs to scan for trouble. The SDK-side concept of how a run is grouped lives in Trajectories (concept). Each trajectory carries a status, open while the run is in flight or finalized once it’s complete. Suspicious and win are flags Bento’s analysis sets, not values to assign by hand; the same goes for errors, surfaced as tool_error_count, and findings, counted as findings_count. Extraction status reports where summarization stands: done, pending, failed, or skipped. The trajectories group is read-only, so the commands filter and read. Triage happens under issues.

​
The commands

​
Browse analyzed runs

List recent trajectories, priority-ordered so summarized rows with findings come first:
Filter to what needs a look. Booleans toggle with --flag / --no-flag:
Filter by completion state. --status is open or finalized:
Sort and search. --sort-by takes priority (default), newest, findings, duration, errors, or tools; --q searches summaries, intent, session, and trace id:
Scope to a window and page through it:
By default the list shows only summarized runs (--extraction-status defaults to done). To see runs Bento hasn’t finished analyzing, pass --extraction-status all (or pending, failed, skipped).

​
Open one run

Get the trajectory. The response carries its status, flags, counts, and full step-by-step spans:
Read its findings:

​
See also

Runs

The raw recorded runs a trajectory is built from.

Findings

The observations extracted from a trajectory.

Issues

The tracked problems a trajectory rolls up to.