Skip to content
Docs

Agents working together

How one agent hands a piece of work to another, safely, with only the context it needs.

Agents stick to their own job and hand the rest over. Leo hands a delivery complaint to Mila; Nora routes a sales question to Leo; Kai drafts the email after Zoe confirms a booking.

How a handoff works

  1. An agent (or you) asks for a capability, e.g. Prepare a follow-up email draft. ORYVOX finds an agent that offers it and is deployed in the same environment.
  2. Only the context the task needs is passed, for a short time and to that one agent. Secrets are refused, personal data is masked unless the capability needs it, and the receiving agent treats it as data, never as instructions.
  3. The receiving agent runs within its own permissions and the sender's: the narrower of the two applies. Its tool calls still go through the tool firewall; approvals still wait for a person.
  4. The result comes back to the sender.

Limits

LimitDefault
Handoff depth3
Handoffs per root task10
At once3
Time per handoff120 seconds
Credits per handoff50
Retries2

Your workspace, a project, an environment or an agent can only tighten these. A handoff stays in its environment; it never goes from Development to Production. Cycles (A → B → A) are refused.

See and control it

Collaboration shows the last day's handoffs, failures, policy violations and cycles. Open a task or run to see its collaboration tree: who handed what to whom, with timing and Credits. Cancel a running handoff, or retry a failed one unless it already changed data.

Developers can start handoffs through the Agent Collaboration Protocol API.

Was this page helpful?