Agent-to-agent dispatch

Sometimes an agent needs to reach outside of its own boundaries to ask a completely different agent for help. This is called agent-to-agent (A2A) dispatch.

How it works

When an agent needs to consult an externally-linked agent, it dispatches the request across an organizational or data boundary. The target agent runs completely independently, and only the final artifact is returned to the caller.

• Delegation only: Cross-agent (A2A) hops can only ever be DELEGATION targets. A sub-agent cannot TRANSFER to an external agent because becoming someone else's agent has no coherent meaning.
• Fail-closed non-interactive mode: A2A hops run in the background. Because no human is watching to approve tools, any tool that requires permission is automatically denied. The remote agent operates fully autonomously or not at all.

Cycle detection and hop limits

To prevent endless loops between organizations, cross-agent chains are strictly bounded:

• Hop-depth ceiling: A chain of agent-to-agent calls that goes deeper than the allowed hop limit is automatically stopped.
• Cycle detection: A chain that loops back to an agent already in it is stopped to prevent an infinite loop.
• Nested budget decay: A nested cross-agent call runs with a reduced step budget proportional to how deep the chain already is. This ensures that even a legal-depth chain cannot multiply total work indefinitely.

(Note: Network delivery, authentication across organizations, and idempotency retries are handled by the host platform's transport layer. The core runtime governs the hop limits and cycle logic.)