The problem
Agent-to-agent delegation turns one call into many, across agents you may not own. Without control, any agent can call any other, arguments go unchecked, and a sub-agent’s response can carry data straight back to the caller. It is the tool problem again, one level up.
How it works
Register the agents your agents may delegate to, and write rules that allow, deny, or require approval for a delegation, scoped by the caller and its tags. Each hop is authorized before it runs, rate-limited, and its response is scanned on the way back.
What you get
- An allow-list of target agents, with allow, deny, or human-approval rules per caller.
- Approval for high-risk delegations, routed to a person before the call runs.
- Rate limits on delegation, so a loop cannot run away.
- Response scanning for secrets and PII on every hop, and an audit record of who delegated to whom.
Pair it with MCP tool governance; see agentic governance.