The runtime authority layer for AI agents.

Access controls decide whether an agent can reach a system. Runtime authority decides whether this specific action, with these parameters, is within what the agent has been trusted to do — at the moment it tries to do it.

The answer is always one of three: ALLOW REQUIRE AUTHORIZATION DENY

Each agent is registered in Authesta and gets its own credential. Every request is tied to the agent that made it, and the organization it belongs to comes from that credential — never from what the request claims.

Credentials can be issued, rotated and revoked at any time, without touching other agents.

An authority profile describes what one agent may do: which actions, under which conditions, and which actions are prohibited outright. Each agent has exactly one active profile, so there is never ambiguity about which rules apply.

An action the profile doesn’t mention is not permitted. A prohibited action is denied explicitly, and the record shows which of the two applied.

When an action needs a person, the approver sees the exact action and parameters the agent requested. Approving it authorizes that action only — not a category of actions, and not a session.

The approval produces a single-use grant that expires. The agent can execute the approved action once; a second attempt, a changed parameter or an expired grant is refused.

Authesta doesn’t only advise. In Authesta Cloud or a Private Gateway, the protected action is executed through Authesta, so a denied action never reaches the target system and an allowed one runs exactly as authorized.

If authority can’t be established — an unknown agent, a revoked credential, a missing or stale policy — the action does not run.

After execution, Authesta compares what was authorized with what was observed — from the target system’s response or an independent lookup, not from the agent’s own report.

MATCH only when every authorized field was observed and matched. DEVIATION when something differs. NOT VERIFIABLE when execution completed but Authesta could not independently confirm every field. Unverified is never shown as a match.

Each protected action leaves one record: the request, the agent, the rule that decided it, who approved it and when, what executed, and the verification result — including where each observed value came from.

Evidence from Private Gateways is synchronized to Authesta Cloud so the full history is in one place.

• Suspend an agent: its new actions are refused.
• Revoke a credential: the next call with it is refused.
• Suspend an organization: all of its agents stop.
• Disable a Private Gateway from the control plane.

Authesta Cloud applies these at once. Private Gateways apply them with their next signed policy update, and stop serving altogether if they can’t get one within a bounded window.

Every one of these is a deliberate, confirmed step, and each is recorded.