What is Decommissioning?
Operations (LLMOps & AgentOps)Retiring an agent deliberately: revoking credentials, archiving decision records, closing the eval set, and answering whether the system should still exist at all. The annual cadence item most operating models skip.
Why It Matters
Decommissioning is the missing half of the agent lifecycle. Provisioning gets the tooling and the attention; retirement gets none, and the result is agent dark matter: systems that no longer serve a purpose but still hold credentials, consume resources, and appear in audit logs. Gartner projects over 40% of agentic AI projects will be cancelled by the end of 2027, which makes retirement a regular operational event rather than a rare one. An operating model without a decommissioning path is an inventory that only grows.
The Six-Step Checklist
Freeze triggers. Disable schedules, webhooks, and event subscriptions so the agent stops receiving new work while cleanup proceeds.
Drain in-flight work. Complete or deliberately cancel running tasks. An agent frozen mid-sequence needs to finish or roll back to a consistent state, not be abandoned mid-write.
Revoke credentials. API keys, service accounts, grants. Per-instance identity makes this one revocation; shared credentials mean finding every one the agent touched.
Remove access. Downstream permissions, integrations, and OAuth scopes that live outside the credential set.
Retain records. Traces, decisions, and eval records archived with a defined retention period. Legal and audit obligations outlive the agent.
Tombstone and verify. Mark the agent deprecated in the registry, then verify no residual permissions, no new traffic, and no orphaned resources remain.
Where It Breaks
Decommissioning research identified the ordering failure: teams delete the agent record first, which destroys the evidence needed to audit what the agent did, and the questions arrive later from regulators, customers, and the security team. Evidence gets retained before the record is marked deprecated, credentials are revoked before access is removed, and triggers are frozen before work is drained. Each step depends on the one before it, which is why it is a checklist rather than a cleanup task.
How Flytebit Handles It
Decommissioning is an annual cadence item in our operating model: every agent gets a “should this still exist” review, and retirement follows the six-step order so the audit trail survives the agent. The full lifecycle is in Operating Agentic AI Systems, and the operating model is our LLMOps work.