Skip to main content

How to retire an AI agent and take back its keys.

A retired agent with live credentials is a cost and a risk. How to decommission an AI agent: the triggers, the steps, and the proof it is gone.

Topic
Agent governance
Product
Novie Control
Format
Guide
Reading time
3 min
In short

To retire an AI agent, stop its schedule and triggers, revoke every credential it holds, remove its access to each system and data source, close its budget, and keep its record. An agent that nobody uses but that still holds keys is a cost and a risk: it can be triggered, its credentials can be stolen, and it keeps permissions nobody is reviewing. Retirement is complete when the inventory shows it gone and the keys are dead.

Why agents outlive their purpose

Agents are cheap to start and nobody's job to stop. A pilot ends and its agent keeps its keys. A team reorganises and the agent's owner moves on. A better agent replaces it and the old one keeps running on a schedule, checking for work that never comes and paying for the check.

The scale is new. Palo Alto Networks' 2026 Identity Security Landscape, a survey of 2,930 security leaders, found machine identities, AI agents among them, now outnumber people 109 to 1, up from 82 to 1 a year earlier. Only 37 percent of organisations could revoke an AI agent's credentials.

Regulators have noticed. Singapore's Model AI Governance Framework for Agentic AI, updated in May 2026, names agent sprawl as a risk. It recommends that every agent identity and its permissions be issued from and tracked by a central system, so an organisation can see what it has deployed and remove identities it no longer needs. APRA's April 2026 letter to industry asks for ownership and accountability across the AI lifecycle, which ends with retirement.

The triggers

  • The purpose is done: the project ended, the pilot closed, the process changed.
  • It was replaced by a newer agent doing the same work.
  • Its owner left or moved, and nobody took it on within a set period.
  • It has not run, or run usefully, for a set period.
  • Its approval expired and nobody renewed it.
  • It breached its scope or its cap, and the decision is to stop rather than fix.

Write the triggers into the approval. An agent approved with a review date and an owner has a natural end; an agent approved without them runs until someone notices.

Seven steps to retire an agent

  1. Confirm the owner and the decision. One person signs that the agent is to be retired, and when.
  2. Stop the triggers. Disable its schedule, webhooks and any event that can start it.
  3. Revoke its credentials. Every API key, token, service account and OAuth grant it holds, at the provider and in every system it reached.
  4. Remove its access. Take it off the roles, groups and data shares it was given. Revoking the key without removing the role leaves a door for the next key.
  5. Close its budget. Set its cap to zero and release its allocation, so any call that slips through is refused.
  6. Hand over its work. Anything that still depends on it moves to its replacement or to a person, with a date.
  7. Keep the record. Its row in the inventory is marked retired, with the date, the approver and what was revoked. The audit trail of what it did stays for as long as your retention rules require.

Proving it is gone

A week after retirement, check three things: no model calls under its identity, no use of its credentials anywhere, and no budget drawn. Any of the three means something was missed. The check is quick when every call and every dollar is recorded against an agent, and slow when it means reading logs from every system the agent touched.

How Novie Control does it

Novie Control is live. Every dollar visible. Every budget governed. Every call optimised. Every agent governed. Every agent carries a named owner, an approval trail and a full audit trail, so retirement is recorded against the same row the approval was. Caps hold before spend is committed, so an agent with a zero cap cannot spend, and every AI dollar is traced to the owner at the moment it is spent, which shows any call made after retirement.

Sources

  1. Palo Alto Networks, Assess maturity when machine identities outnumber humans 109:1 (20 May 2026), from the 2026 Identity Security Landscape
  2. IMDA, Model AI Governance Framework for Agentic AI, Version 1.5 (20 May 2026, updated 5 June 2026)
  3. APRA, Letter to Industry on Artificial Intelligence (30 April 2026)
  4. KPMG, AI Quarterly Pulse Survey Q2 2026 (24 June 2026)

Other articles

Frequently asked questions

What is a zombie AI agent?

An agent that is no longer used or owned but still holds credentials, permissions or a schedule. It can still be triggered, its keys can still be stolen, and it often still spends.

How do you decommission an AI agent?

Confirm the decision with the owner, stop its triggers, revoke every credential, remove its roles and data access, set its budget to zero, hand over its work, and mark it retired in the inventory with what was revoked.

When should an AI agent be retired?

When its purpose is done, it is replaced, its owner leaves and nobody takes it on, it stops being used, its approval expires, or it breaches its scope and fixing it is not worth it.

How do you know an AI agent is fully retired?

A week later, there are no model calls under its identity, no use of its credentials and no budget drawn. Any one of the three means a step was missed.

Build the intelligenceyour organisation compounds.