Runs and evidence
Every claim an agent makes should be checkable in one move from where it is read.
A chat turn records
- Steps — the tree of what the agent did, with confidence and the sources each step cited
- Tool calls — each call's input and output
- Cited documents — numbered, so an inline
[n]opens the source - Tokens and cost — including any specialist the agent delegated to
- Status — complete, incomplete, failed, refused, stopped, truncated or continued, with a reason
- Routing — which agent answered and why
- A Langfuse trace id, where the model and full prompt live when tracing is on
Cards keep who decided them and when.
Runs record the workspace version
Tool calls, workflow runs, mission runs, eval runs and worker runs record the workspace_sha they ran on — the git commit of the workspace plus the base pack and plugin versions — so any output traces back to the exact prompts that produced it.
Where to look
| Page | Shows |
|---|---|
/dashboard/chat/<id> | The conversation, each turn's steps and citations |
/dashboard/activity | One feed of checks, runs, automation fires and syncs (/dashboard/logs redirects here) |
/dashboard/missions | Mission runs and their artifacts |
/dashboard/observability | Traces, cost and latency |
/dashboard/evals | Eval runs against the pass bar |
Turn guarantees
The harness promises every turn ends in a stated outcome: an answer pass, one continuation for an empty or promised-but-not-done turn, one retry after a malformed tool call, a notice when a hand-off failed, and a stated failure or refusal rather than silence.
Hand-off gates
An object type can declare a gate on a transition (for example, a deal becoming proposal): deterministic checks run first, then a judge with the seat's rubric. Below the escalation bar (default 0.6) the work returns to the seat or escalates to a person as a Gate ask. See Object types and gates.
Evals
Per-agent datasets with deterministic checks and an optional judge, scored against a pass bar (default 0.8). See Evals.