Running agents

A run is a single execution of an agent. Runs happen on demand, on a schedule, or from an event — this page covers running on demand and reading the result.

Run an agent now

Open an agent and use Run to execute it once. You can optionally pass an input message; agents with no input run their instructions directly. By default a run uses the agent’s stable version; the chat surface runs the live draft.

The run detail page

Each run records and displays:

• Output — the agent’s final response, rendered as Markdown.
• Status — queued → running → succeeded / failed.
• Tokens & cost — input/output token counts and the computed USD cost, so you can compare models and prompts.
• Tools used — every tool the agent called, in order, with a success/failure mark and (on failure) the error. This is captured for Pydantic agents on both successful and failed runs — so a run that broke before reaching a step still shows what it did call. The same data rolls up in Tool uses.
• Timing & trigger — when it ran and what triggered it (manual, schedule, or event).

When a run fails

Failed runs keep their captured output and tool calls so you can diagnose them. The agent and workspace dashboards group recent failures, and the run page links to related context. Common causes — a missing provider key, an unauthorized or stale connection, or a truncated response — are covered in Troubleshooting.

Improving an agent from a run

If a run is wrong, use Improve the Agent to describe what should change. TAS turns the feedback into a pull request via Tembo and correlates the merged PR back to your submission. See Improvements.