Skip to main content
Connect through remote MCP when supported by your ADK runtime, or define server-side functions that call Darwin’s SDK. Model Search as read-only discovery. Model the seven Act operations separately: start, send, read, list, archive, authenticate, and pay. Persist threadId and the last consumed cursor outside the agent transcript so a later run can recover work with get_thread. Use Darwin’s first-party webLink for connection and payment interactions rather than proxying sensitive provider credentials through the agent.

Runtime state

Keep the selected Search agent, optional nested capability, threadId, last consumed cursor, retry identifiers, and latest Darwin response in session or workflow state. Give the model the current state it needs, but keep secrets and sensitive browser data in the application layer.
Last modified on October 1, 2026