Agent registry, identity and lifecycle: what the 2026 trend says about the AI Control Plane

In most organizations, the first AI agents arrived through the side door: an automation tool configured by a team, an assistant plugged into a mailbox, a coding agent launched by a developer. A year later, the question is no longer "do we have agents?" but "how many, who created them, and with what rights?". Current practice has a name — agent lifecycle governance — and it closely overlaps what TEAF calls the AI Control Plane.

What 2026 practice recommends

Guides published this year converge on a few ideas (Saviynt, AI Journal, EngineerUp):

• An agent is an identity in its own right, with its own credentials, explicit permissions and a documented lifecycle — not an invisible extension of a developer's or user's account.
• A single registry serves as the system of record: which agents exist, who owns them, what data they access, when they were created, when their rights were last reviewed.
• The lifecycle starts before creation and ends after deactivation: identity, permissions, logs and downstream access paths must be retired or archived. This is where many programs become governable — or add a layer of orphaned non-human identities.
• Discovery matters as much as registration: agents nobody declared must also be spotted.

What TEAF's AI Control Plane already covers

The foundation described in the article on the AI Control Plane overlaps several of these points: separation of identity, authentication and authorization for each agent; secrets kept out of code; logged provenance of actions; human validation on critical changes; rollback capability. The explicit boundary between what an agent executes alone and what requires validation is the central governance decision, and it falls to the AI Control Officer, one of TEAF's five roles.

What TEAF does not provide

To be clear: TEAF is a governance framework, not an identity directory product. It does not provide a ready-made agent registry, automatic discovery of undeclared agents, or periodic rights review at fleet scale. Those functions belong to existing identity and governance tools. What TEAF brings is the structure saying what to record and who answers for it: the agent is tied to a capability (Capability component), to a decision that authorized it (Living ADR) and to a named owner.

A minimal agent record

Six fields per agent are enough to start:

• Identifier and purpose: what is it for, in one sentence?
• Owner: a named person, not a team.
• Linked capability and authorizing decision (ADR reference).
• Accessible data and systems, with exact rights.
• Autonomy level: assisted or autonomous execution, and the actions subject to validation.
• Creation date, last review date and planned retirement.

The frozen-inventory trap

A registry filled in once is a document that ages: exactly the problem the Living ADR seeks to avoid for decisions. The registry has value only if kept current by a process (reviews at expiry, alerts on an agent with no owner) and reconciled with what is actually observed (Observation component). The maturity diagnostic helps place yourself honestly.

As the article on the AI Control Plane already says, agent governance remains an open workstream on several dimensions (prompt injection, model drift, agent-to-agent trust). A properly maintained registry is the first stone, not the last.

• An agent is an identity: own credentials, explicit rights, named owner, documented lifecycle.
• A single registry and reviews at expiry prevent orphaned agents from accumulating.
• TEAF defines the structure and responsibilities; it does not provide registry tooling or automatic discovery.
• An inventory not reconciled with observation becomes an aging document again.

Sources

• Saviynt — Managing AI agent lifecycles
• AI Journal — Governing AI agents as enterprise identities
• EngineerUp — AI agent lifecycle management framework for 2026