Learn · Delegation
Multi-hop delegation and agent-to-agent authentication
In multi-agent systems, an agent hands work to a sub-agent, which hands it to another. At the end you need to say which person authorized the chain and what each hop was allowed to do. Without per-agent identity and signed delegation, forged or inflated authority passes down the chain unseen.
What goes wrong across hops
- A sub-agent claims authority it was never given, because messages are plain text.
- Each hop widens what the next may do.
- An approval for a small action is reused for a large one.
- The record shows a chain of calls with no proof of who authorized it.
Controls that hold across hops
- Identity for each agent, checked outside the message text.
- Signed task context bounding what the agent is doing.
- Delegation that can only narrow: a sub-agent never gets more than the delegator holds.
- A single-use approval for one exact operation, bound to its arguments.
- Agent-to-agent checks: one agent verifies another agent's identity.
Attributing the chain
Because the task context, the delegation and the approval hashes are part of each operation proof, the record can show which person approved which agent, which agent delegated to which, and what each operation was.
How Oathera helps
Oathera Agent Enforcement checks the task context, the delegation and the approval on each call, and the audit record keeps the link from action to agent to approving person. You can try the failure modes in the free Multi-Agent Conspiracy chapter at oathera.ai/multi-agent-conspiracy.
Oathera does not verify people, judge model output, inspect or rewrite prompts, or replace your identity provider. Your identity provider and proofing sources do the first; Oathera is the layer under them for agents.
FAQ
How do you authenticate one AI agent to another?
Each agent has its own verified identity, and the receiving agent checks it outside the message text rather than trusting a name in the message.
What is delegation chain attribution?
Being able to show which person authorized which agent, and which agent delegated to which, for every action.
Can a sub-agent get more authority than its parent?
Not with narrowing-only delegation: a sub-agent can only receive a subset of the delegator's authority.
What is a weak approval?
An approval not tied to the exact action, amount or agent, which can be reused for something else.
Related: Enforce least privilege for an AI agent · Audit every action an AI agent takes · Multi-Agent Conspiracy, a free game
Become a design partner Watch the simulator More guides
← Back to Learn