
An agent has no last day
On an employee's last day, the badge, the laptop and the email are switched off. The AI agents they built keep working.
- The access outlives the person. An agent that works through loan files keeps its key, keeps calling the AI model and keeps reading customer records after its builder has gone.
- Nobody answers for it. When an auditor asks who is accountable for that agent, there is no name to give.
- Offboarding misses it. The leaver process was built for people, so nothing in it switches the agent off.
- It is a named risk. The OWASP Top 10 for Agentic Applications 2026 lists Identity and Privilege Abuse as ASI03.¹ IBM found 92% of organizations with an AI-related breach lacked proper AI access controls.²
The question is not whether the agent does good work. It is who answers for it, today, and what happens the day they leave.
No owner, no agent
Argent Strategic Intelligence ties every AI agent that works through it to a person: a real, active employee who answers for it.
- Registration needs an owner. An admin picks the owner from the people list, or gives an email. Leave it blank and the form says: "Pick an owner. Every agent needs a person accountable for it, so a registration without one is refused." Name someone who does not exist, and Argent refuses that too.
- The owner must be real and active. A human, in the same company account, not deactivated, and placed in an org unit or department, so some admin's branch is responsible for them.
- Recorded before it exists. The registration is signed before the agent is created. If the signed record cannot be written, nothing is created.
- An agent always keeps an owner. Ownership can move to a new person, after the admin confirms with their passkey again, but it can never be cleared to nobody.
- Owners can register their own. A person in the owner portal can create agents, and each one is owned by them.
This covers every agent that works through Argent: AI coding agents, customer-facing assistants, enterprise copilots and agents that call tools and other agents.
One question on every request
Argent stands between the agent and the AI model it calls, such as ChatGPT, Claude or Gemini. Registration is not a one-time check: the owner is checked again on every request.
- The agent sends a request. It arrives at the Argent gateway with the agent's own key.
- Argent identifies the agent and looks up its owner.
- Argent asks one question: does this agent still have an active owner, placed in the organization?
- Yes: the work goes through, and the rest of Argent's checks still run.
- No: the request is refused. Missing, deactivated or unplaced owners each have their own reason code (
auth.owner_missing,auth.owner_invalid,auth.owner_unplaced), the agent gets HTTP 403, and nothing reaches the model. - One signed record per decision, allowed or refused: which agent, what Argent checked and what it decided. No sampling.
The check runs on traffic routed through the Argent gateway, which is where agents get their access to AI models. Pair it with keyless agents, so agents never hold a provider key of their own, and agent reach control.
When the owner leaves, the agents stop
The control follows the person's status on its own. Nobody has to remember the agents.
- From your directory. Argent connects to the identity stores you already run: Active Directory and LDAP, Microsoft Entra ID, Okta, Auth0 and Keycloak, and governance tools such as Ping Identity, SailPoint and Saviynt over SCIM. When the directory disables or removes a person, Argent deactivates them and signs the change; a pulled directory (Active Directory and LDAP, Entra ID, Okta, Auth0, Keycloak) also places a hold on their agents, and checks every 30 seconds by default.
- Or by hand. An admin can switch a person off in Argent with the same effect, without waiting for the directory. Once a person is off in Argent, a directory sync cannot turn them back on.
- On the very next request. Their agents are refused at the gateway. Nothing reaches the model.
- Nothing is deleted. Keys, history and records stay. Later, an admin gives the agents a new, active owner and releases them, and they go back to work.
Contain any agent at any time. Admins can Pause (requests refused until resumed; key untouched), Quarantine (every request refused for 1 hour, 24 hours, 7 days or until released; nothing is deleted) or Kill (the key stops working; an admin can later bring the agent back with a new key, once it has an active owner). Each needs a stated reason, and owners can hold their own agents from the portal. The full leaver story is in owner offboarding.
What the operator sees
Every agent has a name, an owner and a status, and the console shows them in plain words, with the reason code beside the English.
| Where | What it shows |
|---|---|
| Agents | Every agent with its owner, scopes and status, and a "Register agent" form where Owner is required. |
| Agent page | The agent's status (ACTIVE, PAUSED, QUARANTINED, KEY REVOKED), its owner with an active or deactivated stamp, department, and the Pause, Quarantine, Kill and Release controls. |
| Person page | A person's status, Disable and Enable, the agents they own, and "This person is deactivated. None of their agents are still active." |
| Org chart | Owners and their agents in place; a refused agent reads "Refused at every request since <name> left" with its 24-hour refusal count. |
| Unplaced owners | Live agents whose owner sits in no unit, so an admin can place the person. |
| Spend by owner | Spend rolled up by owner, with one "No live owner" bucket for the rest. Pairs with AI spend caps. |
| Signed record | Agent, owner present, owner active, owner placed, the checks enforced, and the decision. |
When an examiner asks who answers for an agent, the answer exists before anyone asks. Go deeper with session replay and one-click audit evidence.
Where it fits, and how you start
Identity Tethering is the base of agent accountability: every other control assumes someone answers for the agent.
- Goals: it is the core of agent identity and ownership, and it feeds audit and accountability, compliance evidence, AI visibility and AI budget and token spend.
- Industries: in financial services, insurance, healthcare, the public sector and retail, it gives every agent that touches customer data a named, accountable person. All industries.
- Regulated and air-gapped: it runs inside your own network too. See AI in regulated and air-gapped environments and deployment options.
- Find the agents first with shadow AI discovery, then register them with owners. Watch how they behave with agent behavior watch.
- Watch. With Identity Governance in Watch, Argent records each agent that has no active owner while its requests pass. Agents keep working unchanged. Pause, Quarantine and Kill still hold. See watch before you enforce.
- Assign. Connect your directory (all integrations), give every agent an active, placed owner, then switch on Enforce.
- Prove. Each record names the agent, the owner checks and the decision.