Reading agent runs

Every time an agent does work, that work is preserved as a run — the trigger that started it, the tools it called, the artifacts it produced, the reply it left. Most of the time you don't read runs; you read the inbox summary and click through if you want detail. This article is for the times you do want detail: verifying a result, debugging a failure, or just checking what happened overnight.

Where runs live

Past runs live in the same conversation sidebar that holds your chats. Open any agent → use the filter pills above the list to switch from All or Chats to Runs. The list is grouped by Today, Yesterday, Previous 7 days, and Previous 30 days, with a search box for finding a specific one.

Click any entry to open it.

What a run record shows

Each run carries a small set of facts:

Status. Usually completed (success) or failed (something went wrong). You might also see running for an in-progress one, cancelled if it was stopped, or stopped early if the agent hit its work limit before finishing.
Trigger type. How this particular run started — chat (you asked), task (a manual run on a task), scheduled (a recurring task fired), or team (a captain delegated it).
Started at / completed at. Timestamps and total elapsed time.
Tokens used. A rough cost signal — bigger numbers mean longer or more tool-heavy runs.
Error. Populated only when the run failed.

Reading a successful run

Click the run to open it. It renders as a conversation thread: the original request at the top, then the agent's working steps in order:

The trigger that fired it (manual, scheduled at this time, task X assigned).
Tool calls the agent made — searching the workspace, reading a page, creating a page, posting a comment — with their inputs and outputs. Integration tool calls show up here too, named by connector and action (gmail.search_inbox, linear.create_issue, slack.post_message) so you can audit what the agent touched in each third-party app.
Artifacts produced (pages, tasks) appearing as preview chips you can click to open.
The final reply text.

It's the same view shape as a chat, except you're looking at the agent's full working session rather than the live exchange.

Reading a failed run

Failed runs carry a status of failed and an error message. Common patterns:

Missing tool access. The agent tried to call a tool you didn't enable. Fix: add the tool in the agent's settings.
Permission denied. The agent tried to read a resource it doesn't have access to. Fix: check the agent's project access, or scope the work into a project it can see. If the failure is on an integration call, the connector may have lost auth — reconnect it in Settings → Integrations.
Rate limit or model error. Transient. Try again.
Stopped early. The agent hit a step limit before finishing. Usually means the instructions ask for too much in one run; tighten the brief or split the work across two tasks.

Re-running

There's no "retry this run" button. To run the same work again:

From a task — set Frequency to Run now (or save with the schedule unchanged).
From chat — send the same message again. Each chat message is its own run; the new run is a fresh attempt with the agent's current instructions.

Note that if you've edited the agent's instructions between the failed run and the retry, the new run uses the new instructions. Runs are records of what happened, not snapshots of the agent at the time.

Tips and gotchas

You can't edit a sent message. If the prompt was wrong, send a follow-up rather than trying to rewrite history.
Runs are scoped to the agent. To see who did what across a team, open the team and use the same Runs filter on its conversation sidebar.
Tokens used is a rough cost signal, not a charge — it's there to help you spot runaway runs, not to bill against.