Skip to main content
Your Major apps can start agent runs from their server code and interact with those runs: read their messages, send follow-ups, stop them, and answer their tool approvals. A common pattern is a dashboard that, on load, starts a few agents (summarize my inbox, prep my calendar, review open PRs) and shows their results as they come in. App triggers start agents, not workflows. To start a workflow from your own code, call its webhook.

​
Example requests

  • “When the dashboard loads, have the inbox agent summarize today’s unread email and show the result.”
  • “Add a Research button to each account row that starts the account-research agent on that company.”
  • “Show the agent’s pending approvals in the app so users can approve them without leaving the page.”

​
Triggering from code

Add the agent to your app to get a generated client, then call it from server code:
Runs are asynchronous: run() returns right away and you poll for messages. The client can also send follow-ups, stop runs, and answer tool approvals. See Agents client for setup and the full API.

​
Who the run acts as

When a signed-in user triggers the run through your app, the run is attributed to that user, and agents with per-user connectors act with that user’s credentials. A call made outside a user’s request runs as the person who deployed the app. A call made for an anonymous visitor, such as on a public page, is rejected because there’s no user to attribute the run to. In every case, the deployer must have access to use the agent.

​
Next steps

Agents client

Setup, methods, approvals, and errors.

Webhooks

Start a workflow from your app or backend.

Tool permissions

Decide which tools need approval.