The 2026 stack for AI agents is filling in fast — protocols for how an agent authenticates, frameworks for who it is, and models for what it may do. F-ACT is not another identity protocol. It is the governance envelope that sits above them: a compact, verifiable record of the authority an agent was granted, how it may act, what it did, how that authority is revoked, and how its data is handled.
Identity tells you who the agent is. F-ACT tells you whether it should act — and proves what happened.
The emerging protocols answer who the agent is and how it connects. F-ACT answers whether it should act — and leaves proof.
Defines how an agent authenticates to a tool or data server and carries scoped authorization tokens between them.
Learn moreMCP moves the token; F-ACT governs whether that grant should exist. F-ACT records who authorised the agent and to what scope, and makes the grant revocable and auditable.
Uses Decentralized Identifiers (DIDs) and Verifiable Credentials so an agent — and the human principal behind it — can be cryptographically identified across organisations without prior coordination.
Learn moreF-ACT sits on top of a verified identity: once you know who the agent is, F-ACT is the record of what it is allowed to do and what it did.
Applies Zero-Trust principles to autonomous agents with an “AND-gate”: an agent’s effective permission is the intersection of its own authorized scope and the delegating human’s permission.
Learn moreF-ACT’s Grant and Usage dimensions express exactly this least-privilege, deny-by-default posture — but as a portable, verifiable claim you can publish and check, not only a runtime policy.
Issues purpose-built agent identities from identity blueprints, so agents do not hold their own long-lived keys.
Learn moreWhere Entra provisions the identity inside one cloud, F-ACT provides the vendor-neutral, portable conformance record that travels with the agent across clouds, vendors and organisations.
Early US government framing of controls and reference architectures for autonomous AI agents.
Learn moreF-ACT gives an implementable, testable expression of the agent-level controls such guidance calls for. See the regulatory crosswalk for the mapping to formal regimes.
Agents increasingly hand work to other agents. Each hand-off is a delegation of authority, and each one is a place where authority can silently grow. T-RUE and the OCC Agent Record treat a delegation chain as part of the agent’s declared authority.
Each hop passes on the same or less authority than it holds. An agent can never grant a downstream agent more than its own effective authority or its ceiling.
The originating principal, every intermediate agent and the scope passed at each hop are recorded, so any action can be traced back to the person or organisation that authorised it.
Every agent in the chain uses its own identity. Tokens are exchanged or attenuated per hop (for example OAuth 2.0 Token Exchange, RFC 8693), not forwarded as-is.
Revoking a grant at any hop withdraws every delegation derived from it. T-RUE’s revocation route is tested at the top of the chain and at least one downstream hop.
The record states the maximum delegation depth. Unbounded re-delegation is treated as a Rogue Agent risk (OWASP ASI10).
A chain is classified at the highest AIH-12 horizon of any agent in it. Low-horizon agents do not dilute a high-horizon one.
| Hop | Who | Scope passed on |
|---|---|---|
| Principal | Finance team, Example Pty Ltd | Pay approved supplier invoices up to A$10,000 each, until 31 Dec 2026 |
| Hop 1 | Accounts-payable agent (H5) | Pay approved invoices up to A$10,000; may ask a lookup agent for supplier data |
| Hop 2 | Supplier-lookup agent (H3) | Read-only access to the supplier directory. No payment authority passed on. |
Illustrative example. Authority narrows at every hop; revoking the principal’s grant withdraws both agents’ authority.
The NIST National Cybersecurity Center of Excellence (NCCoE) published a concept paper, Accelerating the Adoption of Software and AI Agent Identity and Authorization, on 5 February 2026; public comments closed on 2 April 2026. It proposes showing how OAuth 2.0, OpenID Connect, SPIFFE/SPIRE and Next Generation Access Control (NGAC) can be applied to agents. OCC tracks that work and aligns the Agent Record to it; where NIST guidance is final, it prevails. OCC is not affiliated with or endorsed by NIST.
Each agent has its own non-human identity, distinct from the person or service that launched it.
Agent Record: agent.identifier — a stable identifier (e.g. a SPIFFE ID, OAuth client ID or DID). Never a shared human account.
Agents authenticate with short-lived, verifiable credentials rather than long-lived secrets.
Agent Record: Evidence entry against the identity control, with the credential type and lifetime. Digest the configuration if it can be shared.
Permissions are scoped to the task and enforced at the resource, with least privilege.
Agent Record: authority.effective lists what the agent can actually do; authority.declared what it is meant to do; reconciled_on shows when they were last compared.
Where an agent acts for a person or another agent, the chain of authority is explicit and attenuated at each hop.
Agent Record: principal names who the agent acts for; declare each downstream agent and the scope passed to it (see multi-hop delegation).
Actions are logged and attributable to the agent and its principal.
Agent Record: F-ACT Audit dimension; evidence entries with digests and attestation levels A1–A3.
Credentials and delegations can be withdrawn quickly when something goes wrong.
Agent Record: revocation.endpoint and revocation.tested (T-RUE). An untested route is recorded as a gap.
These are independent efforts by other organisations, described here from their public materials. Listing them implies no endorsement in either direction, and the details of fast-moving drafts may change.