A cloud agent is a one-shot or scheduled task that runs the assistant outside of an interactive chat. You give it a prompt, pick sources, and walk away. It runs to completion, then shows up in the cloud agents page when it's done.
Use this when:
- You want the model to do something long-running (research, multi-step analysis) without holding the chat open.
- You want the same prompt to run every morning or every Monday (a digest, a status report, a refresh of cached numbers).
Where it lives
Cloud Agents in the sidebar app-nav, or /cloud-agents directly. The page has two sections:
- Agents: recent one-shot runs.
- Automations: scheduled workflows.
If anything is currently running, a Running banner sits at the top with live status.
Dispatching a one-shot agent
Click New Agent in the Agents section. The dispatch dialog asks for:
- Prompt: what you want the agent to do.
- Model: defaults to your current chat model; override for heavier or lighter runs.
- Sources: pick any sources from your library that the agent should be able to read.
- Web Search: toggle on if the agent needs current web info (only available if your org has web search enabled).
- Advanced (collapsed by default): max iterations and timeout.
Pick Run mode: Now, hit Dispatch, and the agent kicks off in the background. The row appears under Running. When it finishes, the result moves into Recent with the full output and any artifacts produced.
Dispatching from chat
You don't have to leave the chat to start an agent. The composer's actions menu has an Agent option that opens the same dispatch dialog pre-filled with your current message and the conversation's attached sources. Handy when a question turns out to be too long-running for an interactive turn.
Scheduled automations
Pick Run mode: Schedule in the dispatch dialog instead of Now to save it as an automation. A schedule picker appears where you set:
- Cron expression: e.g.
0 9 * * *for daily at 9am. - Timezone: defaults to your browser's timezone.
The automation starts running on that schedule immediately. Each run produces an agent task you can review on the same page.
To edit, delete, or build more complex multi-step automations, click an automation card. The automation builder lets you chain agent steps, add triggers and conditions, and review run history.
Reviewing results
Click any finished agent row to see:
- The full prompt and final answer.
- Sources the agent actually used.
- Token usage and runtime.
If an agent fails or hits its iteration cap, the row shows the failure reason and the partial transcript.
Where next
- Files and sources: picking what the agent can read
- Chat: interactive use, including the "Agent" shortcut from the composer
