Agent-to-agent delegation is the handoff of work from one AI agent to another, often across different tools or identity contexts. It expands the governance boundary because the original actor no longer controls every action, and inherited permissions can create risk that the first approval never covered.
What Agent-to-Agent Delegation Means in Practice
Agent-to-agent delegation is not just a routing choice, it is a transfer of authority. The first agent hands work to another agent, but the original requester may no longer directly control the downstream actions, tool calls, or environment context that the second agent uses.
That makes delegation different from simple message passing. Once work crosses an agent boundary, the important question is no longer only what task is being asked, but what authority, constraints, and evidence of approval are being inherited along the way.
Why Delegation Chains Create Governance Boundaries
Delegation becomes harder to govern when one agent can hand off to another across different tools, services, or trust domains. Each hop can widen the blast radius if the next agent receives more access than the original approval intended.
In multi-agent systems, this is especially important because the initiating agent may not see every subordinate action. A design that looks safe at the first approval step can still allow a later agent to act with broader scope, different ambient credentials, or assumptions that were never explicitly reviewed.
That is why delegation should be treated as a control boundary, not just an implementation detail. The security problem is not delegation itself, but unconstrained inheritance of privilege and context across the chain.
Identity, Authority, and Context Inherited by Agents
Delegated authority has to answer who is acting, on whose behalf, and with what limits. In agentic systems, those questions often map to token exchange, scoped credentials, task-bound claims, or other mechanisms that let a downstream agent prove what it is allowed to do.
The most common failure mode is ambiguity between delegation and impersonation. If an agent receives a reusable bearer token or an over-broad session context, it may appear to be acting for the original principal while actually operating with wider reach than intended.
That is why good delegation models narrow authority at each step. The delegated agent should receive only the minimum access needed for the specific task, and that authority should be auditable, time-bound, and revocable.
For background on how AI agents establish and carry identity across handoffs, see Agentic AI Identity Guide, and for the standards basis for token exchange and on-behalf-of flows, see RFC 8693: OAuth 2.0 Token Exchange.
Where a delegation chain spans multiple agents and approvals, the practical control question is whether each hop preserves intent or silently expands authority. That is the difference between controlled delegation and privilege drift.
Containment, Revocation, and Observability Across the Chain
Delegation only stays safe if the organisation can contain it, revoke it, and reconstruct it later. If a downstream agent misbehaves, inherited access must be cut off without waiting for the original requester to be involved again.
Visibility matters because the security boundary often shifts during execution. A chain that looks harmless in design review can become risky in production if logs do not show which agent acted, which context it inherited, and which tools or resources it touched.
That is why delegated agent systems need strong traceability around approval lineage, token use, and action history. Without that evidence, the organisation cannot tell whether the work completed as intended or whether the delegated authority was misused somewhere in the chain.
For a deeper treatment of multi-hop delegation, containment, and chained authority, see Multi-Agent and A2A Security Guide. For delegated access revocation and incident containment, see Revoking Delegated Access During an Active Agent Incident.
Risk and Threat Considerations
Agent-to-agent delegation creates a real risk of privilege amplification, confused-deputy behaviour, and incomplete revocation. If a downstream agent inherits broad credentials or implicit trust, an attacker who compromises one agent can often pivot through the delegation chain and reach actions that were never meant to be directly available.
Failure mechanism: Delegated authority is passed without tight scoping, audience binding, or hop-by-hop checks, so the receiving agent can act with more power than the original approval covered.
Impact: A single compromised or over-trusted agent can trigger unauthorized tool use, data exposure, lateral movement across services, or persistent access that survives the original request boundary.
Standards & Framework Alignment
This section maps relevant standards and security frameworks to the operational risks and controls described in this guidance.
OWASP Agentic AI Top 10 addresses the attack and risk surface, while NIST SP 800-53 Rev 5 sets the governance and control requirements practitioners need to meet.
| Framework | Control / Reference | Relevance |
|---|---|---|
| OWASP Agentic AI Top 10 | ASI03 — Identity & Privilege Abuse | Agent delegation is a core identity and privilege boundary in multi-agent systems. |
| Recommendation — Enforce per-hop authorization and task-scoped delegated access for agent handoffs. | ||
| NIST SP 800-53 Rev 5 | IA-9 — Service Identification and Authentication | Agent-to-agent delegation depends on service or workload authentication across trust boundaries. |
| AC-6 — Least Privilege | Delegated authority should be narrowed so downstream agents receive only needed access. | |
| AU-2 — Event Logging | Delegation chains need traceable records of who acted and what was inherited. | |
| Recommendation — Authenticate each agent-to-agent hop before any delegated action is accepted. Limit delegated agent permissions to the minimum required for the task. Log delegation events, inherited context, and downstream actions for review. | ||
Practitioner Guidance
Why practitioners should care: The governance question is not whether agents can collaborate, but whether each delegation step is deliberate, least-privileged, and reversible. If the system cannot show what authority was handed off, the organisation cannot confidently approve the chain.
Practitioner takeaway: Treat every agent handoff as a new authorization event, not a continuation of the old one, and make the delegated scope smaller than the originating agent's own standing capability.
Related resources from NHI Mgmt Group
Deepen Your Knowledge
Free weekly newsletter
Subscribe to the NHI & AI Identity Journal
The latest on NHI and Agentic AI security – articles, research, breaches, news and events every week.
Bonus 33% off our NHI Course when you subscribe.
Reviewed and updated by the NHIMG editorial team on October 31, 2026.
NHI Mgmt Group — the #1 independent authority on Non-Human Identity, IAM, and Agentic AI security. nhimg.org