LEARN · AI AGENT AUTHORITY
AI Agent Authority: How It Relates to Identity and Authorization
AI agent authority is a current, explicit mandate for an authenticated agent to perform a particular action on a particular resource under stated constraints.
CORE DISTINCTION
Identity tells you who.
Authority explains why this action is allowed.
DEFINITION
Authority answers a narrower question than identity.
An agent can authenticate, discover a tool and decide to use it without holding organisational permission for its consequence. VAGP is the public protocol that models that permission as verifiable, revocable mandate authority. Verimand is an implementation and product ecosystem built around that separation.
Who is acting?
Authentication binds the request to an agent. A valid identity alone does not grant a mandate.
What can it technically do?
Tools and attested capabilities describe ability. Tool availability and attestation do not create authority.
What does it propose to do?
A model or decision system can select an action. Its confidence cannot authorize that action.
Why may it act now?
A current, explicit mandate must cover this agent, action, resource and applicable constraints.
MANDATES
Only a mandate creates authority.
A VAGP 0.3 mandate binds an issuer and subject to allowed actions, resources, constraints, validity and the agent’s exact attested Agent DNA state. Delegation can preserve or narrow these bounds; it cannot expand them. Two incomplete authority paths cannot be pooled into a new permission.
IAM, OAuth and policy engines remain useful technical controls. VAGP adds a separate mandate and evidence layer for the agent’s particular action; it does not replace those controls.
Read the public VAGP 0.3 specificationACTION BOUNDARY
Authority must still be current when the action executes.
A past approval is insufficient if the mandate was revoked, trusted Agent DNA state changed, status evidence became stale or a permit was already consumed. VAGP resolution and verification fail closed when required authority cannot be established.
A qualifying path yields a single-use Execution Permit bound to the exact request. The enforcement boundary verifies and consumes that permit before submitting the protected action.
One CloudOps example.
The same authenticated agent can inspect a sandbox workload but cannot modify configuration outside its mandate. In the controlled Gateway pilot, the out-of-scope request was denied before the protected tool executed.
Follow the execution-time chainEVIDENCE AND LIMITS
Evidence must not claim more than was observed.
VAGP separates authority decisions, execution permits, and signed INTENT and OUTCOME evidence. Evidence Certainty preserves uncertainty: a provider timeout is not proof of success or failure. A witness records what happened; it is not another source of authority.
Public VAGP 0.3 Core conformance and an independent Go implementation support protocol implementability. Verimand’s controlled Gateway and real multicloud validations have narrower scopes. They do not establish production SaaS readiness or validate every identity provider.
Inspect validation and limitsREAD NEXT
Continue from concept to boundary.
The next guides cover the execution path and the separate roles of agent-to-tool and agent-to-agent protocols. The frozen VAGP 0.2 reference remains available for its earlier, explicitly versioned cloud proof.