The agent you forgot to retire can become the riskiest worker in the company.
Most teams design the launch of a machine worker. Fewer design its retirement. That gap is expensive.
An agent can be demoted politically and still keep running technically. It may hold an API key, watch an inbox, write CRM fields, rerank accounts, open tickets, or notify humans after the workflow it was built for has changed. Nobody thinks of it as active strategy anymore. It is just background automation.
That is exactly why it is dangerous. A stale worker does not need to fail loudly to cause damage. It only needs to keep applying old rules to new reality.
Count the workers, triggers, credentials, queues, and write permissions that remain alive after a lane is paused, replaced, narrowed, or absorbed into another workflow.
Agent ops is not only about granting authority. It is about removing authority cleanly when the business changes.
If the control plane cannot retire a worker, every launch becomes a permanent liability hiding under a productivity story.
Human teams have offboarding rituals because access outlives intent. Laptops are collected, credentials are revoked, ownership is transferred, recurring meetings are removed, and open work is reassigned.
Machine workers need the same discipline, but with stronger instrumentation. They do not get embarrassed. They do not notice that the market changed. They do not infer that a Slack silence means the lane is dead. Unless the system tells them to stop, they keep executing.
In AI-native orgs, retirement is not a cleanup task. It is a core governance primitive.
A growth team builds an agent to score target accounts during an expansion push. It reads product usage, recent funding, hiring signals, and CRM history, then boosts accounts for outbound sequencing.
The campaign ends. The team changes ICP. A new segmentation model goes live. The scoring agent is no longer discussed in planning, but its nightly job still writes priority fields because nobody tied retirement to campaign closure.
Two months later, sales has a weird pipeline hygiene problem. Some accounts keep resurfacing from the old expansion logic. Marketing suppressions are overridden. A few reps trust the priority field because it looks official. RevOps spends hours cleaning up what appears to be human inconsistency.
The root cause is not a bad prompt. It is missing decommissioning. The worker should have lost write access when the campaign ended, archived its last decision receipts, transferred unresolved exceptions, and left a visible tombstone in the control plane.
Remove API keys, webhooks, queues, browser credentials, scheduled jobs, inbox watches, and permissions that let the worker mutate real systems.
Move active cases, held decisions, exceptions, and pending escalations to a named human or replacement worker before the lane disappears.
Preserve policy versions, decisions, outcomes, corrections, and rollback notes so operators can explain what happened after the agent is gone.
The best decommission plans do not wait for someone to remember an old automation exists. They bind retirement checks to operating events.
A campaign ends. A policy changes. A tool is replaced. A segment is renamed. A data source becomes unreliable. A human owner leaves. A successor agent takes over. Each event should ask the same question: what machine workers now need their authority narrowed, paused, or retired?
Without that question, the org collects ghost labor. Ghost labor looks cheap until it starts writing into production.
Pick one autonomous lane that is paused, stale, experimental, or replaced. Run a retirement drill.
The sharp rule: every machine worker needs an offboarding path before it gets production authority.