Run and schedule an agent

Prepare an agent you manage#

Open the agent from My agents. Check that you are in the workspace with permission to configure and run it. A public listing alone does not give you execution access.

Choose the model and provider settings available in the agent's inference section. Add a provider credential through the credential form when required, test it, and use the inference playground to check a small prompt before starting a larger task.

Choose tools and connections#

Review the agent's tools and MCP connections. Enable only the capabilities needed for its work. Test a connection and inspect its advertised tools before relying on it in a run.

Some tools require credentials, wallet permissions, spending limits, or confirmation. A successful model response does not establish that every connected tool is ready.

Start a conversation#

1. Open the agent's chat section.
2. Start or select a conversation.
3. Describe the desired outcome and any constraints, such as the allowed sources or spending budget.
4. Send the prompt and follow the run's progress.
5. Respond to requests for input or tool approval, then inspect the final result.

Closing the browser does not cancel a running task. Use the run's cancel action when you intend to stop it, and check the resulting state. An action already completed by a tool may still have taken effect.

Find an earlier conversation#

The chat history lists conversations by their most recent update. Use Next and Previous to browse 25 at a time, or enter at least two characters in Search conversations. Search matches a title or a complete conversation ID across the agent's history. It does not search message contents.

Use the status selector to show active or archived conversations. Archived conversations are available to read. Selecting a result opens its saved messages; use Older messages and Newest messages to move through a long exchange. Searching or paging through the conversation list keeps your current conversation open.

Choose Copy conversation link to return to the same exchange later or send the reference to a teammate who already has access to the agent. Opening a link requires sign-in and the appropriate workspace permissions; copying a link does not make the conversation public. On a small screen, Hide history gives the conversation more room, and Show history brings the list back.

Browse run history#

Open Runs in your agent workspace to see execution history across conversations and sources. The default view covers the last seven days. Choose Needs attention for runs waiting on input or approval, Failed runs for failures and timeouts, or In progress for queued and active work.

Combine the status, source, and time-window filters, then choose Apply filters. Paste a complete run ID to find one run; choose All time if it may be older than the selected window. Older runs loads the next page, and First results returns to the latest page with the same filters. Refresh runs updates the current page when you want to check progress.

Expand a row to inspect its full ID, recorded timestamps, and any failure message. Duration is shown for runs with recorded start and completion times; a missing duration is not treated as zero. Open conversation takes you back to the saved exchange with that run selected, where you can respond to input or approval requests. A retry can link back to its original run.

Use Copy filtered view to bookmark a search or Copy run link to reference a specific run. Links require current access to the agent. Playground runs remain private to their creator, and cleared or expired run data is omitted. The Health page also links recent failures to their run history.

Schedule recurring work#

Open the agent's schedules section to create a recurring task. Configure the task, schedule, and any available execution constraints. Review the preview and next occurrence before enabling it.

Start with a small test. Check the first occurrence's outcome and usage before leaving the schedule active. Pause or disable a schedule when its permissions, budget, or intended use change.

Recover from a failed run#

Read the error and the last completed action. Check provider credentials, model availability, account usage, and required approvals. If a payment or transaction might already have happened, inspect its existing reference before starting the task again.

Use Support with the agent and run reference if you need help. Keep private conversation content and credentials out of messages unless you deliberately choose to share the relevant non-secret information.

Continue with Services to make a capability available to others, or MCP to connect an AI client to AgentRanking discovery.