When to retire an AI agent: five signals and a clean shutdown

Retire an AI agent when one of five things is true: the job it was built for no longer exists, nobody acts on its output, a larger agent now does the same work, the model or system underneath it is being retired and the rewrite costs more than the agent returns, or its exception rate has climbed to the point where a person is doing the job again. Retirement is a scheduled action with a date, an owner and a checklist. A quiet switch-off is not retirement, it is abandonment with the credentials still live.

This post is the retirement half of a pair. The maintenance half covers who keeps an agent alive. Here: how to decide an agent’s time is up, what is not a reason, and the shutdown steps that leave nothing behind.

Key Takeaways

• Gartner predicted in June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027; retirement is a normal outcome, plan for it
• Five signals justify retirement: job gone, output ignored, work absorbed by a bigger agent, dependency retired, exception rate too high
• A newer model, one bad week, or a new owner who does not understand the agent are not reasons
• Shutdown is six steps: announce, freeze, redirect, revoke, archive, delete after a recovery window

Why retiring agents is part of running them

Agents get retired for the same reasons software gets retired, only faster, because the ground under them moves on a published calendar. Gartner’s June 2025 press release put a number on it: over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. The same release estimated that only about 130 of the thousands of vendors selling “agents” offered real agentic capability, a pattern Gartner called agent washing (Gartner, June 2025).

The count is about to get large. In April 2026 Gartner forecast that the average global Fortune 500 company will run more than 150,000 AI agents by 2028, up from fewer than 15 in 2025, while only 13% of organizations believed they had the right agent governance in place. Of its six steps for managing agent sprawl, the third is to define agent identity, permissions and life cycle, including “continually reviewing and retiring redundant agents” (Gartner, April 2026). Retirement, in other words, is a governance step, not a failure.

S&P Global Market Intelligence’s 2025 Voice of the Enterprise survey found that 42% of companies had abandoned most of their AI initiatives, up from 17% a year earlier, and that the average organization scrapped 46% of its proofs of concept before production (S&P Global, March 2025). Those were abandoned pilots. Retiring a production agent is different: it has consumers, credentials and logs, and all three need handling.

Our own register shows the shape. As of 30 June 2026 the fleet behind this site had 41 registered agents, 9 live and 5 in beta. The gap between registered and live is not all work in progress. Some entries are agents we retired, kept in the register with a retired mark and a note saying where the job went, because a deleted row cannot answer “did we ever try this”.

The five signals that an agent should be retired

A signal is a measured condition you can state in one sentence. If you cannot write the sentence, you do not have a signal yet, you have a feeling.

Signal	What you can measure	Usual decision
The job is gone	Zero inputs for 30 or more days; the process it served was changed or closed	Retire
Output is ignored	Produced items not opened or acted on for 3 consecutive cycles	Retire or re-scope
Absorbed by a bigger agent	Two agents produce the same artifact for the same reader	Fold into one, retire the other
Dependency retired	Model or API shutdown date published; rewrite cost exceeds 12 months of value	Migrate or retire
Exception rate too high	Hand-backs to a person exceed the agent’s completed runs for a full cycle	Fix within one cycle, else retire

The job is gone. A supplier stopped sending the file, the marketplace closed the program, the report’s only reader left the company. The agent keeps running on empty and keeps billing model calls. This is the easiest call and the most often missed, because an agent with no input produces no errors.

Output is ignored. Our operations cockpit tracks what each agent produced, what was accepted and what was quietly ignored. The third column is the retirement list. If three consecutive cycles of output were never opened, the agent is not doing work, it is generating files. Ask the intended reader one question in writing: “Would you notice if this stopped?” A no is the decision.

Absorbed by a bigger agent. Fleets grow by adding one agent per job, then consolidate. Two cases from our own fleet: a single-report marketing review agent was folded into the marketing chief-of-staff agent as one weekly routine, and a reorder-reminder agent was folded into the customer-reactivation agent’s lane. In each case one register entry was retired and the job name moved to the surviving manifest. A third case was not an agent at all: a nightly data sync with no judgment in it, reclassified as a scheduled job and removed from the agent register.

The dependency is being retired. Model vendors publish shutdown dates. Anthropic’s deprecation page, checked 5 October 2026, lists Claude Sonnet 4.5 as deprecated on 30 September 2026 with retirement on 30 November 2026, and now publishes “not sooner than” retirement floors for active models. It also notes that cloud resellers set their own schedules (Anthropic model deprecations).

Google’s Gemini API deprecations page, updated 1 October 2026, says its listed dates are the earliest a model might be retired, with the exact date communicated in advance (Google, October 2026). When a notice lands, price the migration. If the rewrite costs more than the agent returns in a year, retire it with the model.

The exception rate is too high. Every agent hands some cases back to a person. When hand-backs exceed completed runs for a full cycle, the person is doing the job with extra steps. One repair cycle is fair. Two is a retirement.

What is not a reason to retire an agent

A newer model exists. Model upgrades are maintenance, not retirement, and they can break the prompts and checks built around the old model. Upgrade on your own schedule, test against the saved run log, and keep the agent.

One bad week. A supplier changed a format, an API returned errors for two days, the agent produced junk. That is an incident. Fix, replay, move on. Retire for a trend, never for a spike.

A new owner does not understand it. That is a handover failure. The fix is the handover document and a recorded walkthrough, not a shutdown.

The per-run cost looks high. Compare it to the person-hours it replaced, at a loaded rate, over the same period. An agent that costs a few dollars a day and saves a controller two hours a week is cheap.

The six-step retirement checklist

The NIST AI Risk Management Framework Playbook, under GOVERN 1.7, asks organizations to have “processes and procedures in place for decommissioning and phasing out of AI systems safely”, covering downstream dependencies, data retention, migration to a replacement, and “where and for how long decommissioned systems, models and related artifacts are stored” (NIST AI RMF Playbook). That is the standard. Here is the procedure we run.

1. Announce the date in writing. Tell every person who receives the agent’s output, two cycles ahead, what stops, when, and where the job goes next. Silence here is how a retired agent gets rebuilt in a spreadsheet by someone who never got the memo.
2. Freeze, do not delete. Disable the schedule and the triggers. The code, prompts and configuration stay exactly where they are. A frozen agent can be unfrozen in a minute. A deleted one cannot.
3. Redirect the consumers. If a successor agent took the job, point every reader, inbox and dashboard at it before the freeze date, and confirm the first successor output arrived.
4. Revoke the credentials. Every API key, service account and webhook secret the agent held gets rotated or removed on the freeze date. Palo Alto Networks’ 2026 Identity Security Landscape, a survey of 2,930 IT and security leaders published in May 2026, counts machine identities outnumbering humans 109 to 1, with 9 in 10 organizations reporting an identity-related breach in the prior 12 months (Palo Alto Networks, May 2026). A retired agent’s live key is a machine identity nobody is watching.
5. Archive the artifacts. Run logs, the final prompt set, the manifest and the last 90 days of output go to cold storage with a retention date. Mark the register entry retired, with the date and the successor’s name. Never delete the row.
6. Delete after a recovery window. Microsoft’s Agent 365 governance documentation, dated 23 September 2026, deletes agents with a 30-day recovery window before permanent removal, and includes a step for reassigning agents that have no owner (Microsoft Learn, September 2026). Use the same shape: 30 days frozen, then remove the runtime.

An agent that passes all six steps leaves behind a register entry, an archive and a successor. Nothing else.

Who makes the call

Someone has to own the decision, on a cadence, with the numbers in front of them. On the Fractional AI Officer retainer ($2,900 a month, cancel monthly) that review is part of the weekly written brief: which agents ran, which produced output nobody used, which have a model notice pending, and which we recommend retiring. One build ships every month, and some months the build is a retirement, because a fleet that only grows is a fleet nobody is reading.

Each agent we build is a tomte, our word for one production agent with one defined job. The job definition is what makes retirement clean: when the job ends, so does the tomte, and the register says so.

Start async

If you have agents running that nobody has reviewed since launch, list them at /start. You get a written reply within one business day, no meetings, naming which ones we would check for the five signals first.