Often, and on purpose. Across MING Labs’ own production fleet, 15 to 20 percent of agents are decommissioned within their first quarter. Three retirements carry names: Major Tom, CHRISTINE, and SM3CB v1. The fleet has run in production since January 2026, each agent under a named human owner, and the rate is treated as an operating cost of hiring agents honestly, not as a failure of the program.[S1]
01The rate, and why it is not a failure
Nobody publishes their agent retirement numbers, which creates the impression that healthy fleets keep everything running. The wider data says otherwise. Gartner projects that over 40 percent of agentic AI projects will be canceled by the end of 2027, driven by escalating costs, unclear business value, and inadequate risk controls.[S3] In Camunda’s survey of 1,150 IT and business decision makers, 71 percent of organisations said they use AI agents, yet only 11 percent of agentic AI use cases had reached production in the past year.[S4] Shutdowns are not the exception in this market. Unreported shutdowns are the norm.
The takeaway: the decommissioning rate itself does not tell you whether a fleet is healthy. What produced the rate does.
| Observed pattern | What it usually means | What to check |
|---|---|---|
| Zero retirements, ever | Nobody is measuring use, so nothing ever fails | Whether any agent has a defined shutdown criterion |
| 15 to 20 percent retired in the first quarter | A shutdown criterion exists and is applied early | That retirements are decided by the named owner, on the criterion |
| Cancellation at budget review | The shutdown happened by exhaustion, not by decision | The Gartner pattern: cost and unclear value, discovered late |
02Three retirements, by name
MING Labs keeps its retirements on the record, because a fleet’s failures are data. Major Tom, the first agent, was shut down after 13 days and 764 messages because not one output was ever used or acted on.[S2] CHRISTINE was retired. SM3CB was retired in its first version and rebuilt; its successor runs in the fleet today.[S1]
What the three have in common is who decided and on what. Each retirement was a role decision made by the named human owner against a use criterion, not a verdict on the underlying model. The model that powered Major Tom powers agents that work. The role, coordination without a domain and without accountability, was the thing that failed, and no amount of model improvement would have fixed it.
03What has to be true before you can switch one off
A retirement is only cheap if the organisation can absorb it. Two conditions make that true. First, the work needs somewhere to go: back to the named human owner, or into a redesigned role that gets rehired the way Major Tom’s successors were. Second, the owner must still be able to do the work unaided, which is what the Comprehension Obligation tests every quarter. An agent the team cannot switch off without losing the capability is not an asset. It is a dependency with a personality.
The criterion for the decision itself, the Acted-On Rate, and the full thirteen-day case behind it are on the companion page: when should you shut down an AI agent .
04The measurement, scoped
The number: 15 to 20 percent of agents decommissioned within their first quarter, three retirements by name.
Measurement scope
- Period: January 2026 to July 2026, MING Labs’ hybrid operating model in production.
- Sample: MING Labs’ own agent fleet, agents hired into named roles under named human owners.
- Measured: permanent retirement of an agent role within its first quarter of production, decided by the named owner.
- Counted as retirement: Major Tom (March 2026), CHRISTINE, SM3CB v1. A rebuilt successor in the same role, as with SM3CB, counts as one retirement plus one rehire, not as a survival.
- Not claimed: an industry benchmark. This is one company’s operating record; external figures on the page are Gartner’s projection and Camunda’s survey, cited separately.
- Disclosure: scope only. Internal shutdown protocols and per-agent review records are not published.
Retirement is one half of the same discipline that governs promotion. Agents in the fleet earn autonomy levels as their output gets used, and lose them, or their role, when it does not. That is what it means to run agents inside a hybrid organisation : colleagues are hired, reviewed, promoted, and sometimes let go, and the org chart stays honest at every step.
