The Agent Estate Standard is published today. One requirement in it cannot be satisfied by writing anything down.
Ten issues. I have read every reply myself.
The questions I asked were narrow ones. Do you have a scheduled check comparing agent output against a baseline? Could you produce a complete list of your production agents within forty-eight hours? Reply with one word.
Almost nobody replies with one word. People answer the question and then keep going, and what comes after the answer is the part I have kept. Agents that had drifted for months before anyone noticed. A cost per outcome that no existing method could produce. A register that turned out to be a list of unknown completeness.
I have never asked anyone about retirement. And in a year of people telling me, unprompted, what is going wrong inside their organisations, not one has told me about an agent they switched off.
That is the finding. It is also why one requirement in the Standard I have published today is the one I expect almost nobody to meet.
7.3.7. The organisation must be able to evidence at least one agent retired or demoted as a result of an estate review decision within the preceding twelve months.
Everything else in the document can be written. This one has to have happened.
Nobody has written to me about it. I have seen it from the inside exactly once.
A telco I worked with had seven agents in production and a board report showing 4.1 million dollars in annual savings. Healthy numbers, growing footprint, the kind of slide that earns the next round of funding.
Their new CFO asked one question nobody had asked before. Which of these seven would survive a unit economics review today?
Nobody had ever run it agent by agent, so they did. Three of the seven were destroying value. The aggregate had covered it, because four good agents will carry three bad ones for as long as anyone only looks at the total.
They retired the three. The board number fell to 3.4 million, and for the first time it was true.
The way I usually tell that story, the numbers carry it. Reading it again while writing this Standard, a different detail is the one that matters.
Seven agents had gone into production. The estate had been running for years. And the question that ended three of them came from someone who had been in the building a matter of weeks.
Nobody in that story was negligent. Every person involved was doing their job correctly. What the estate lacked was anyone whose job it was to ask, so the question had to arrive from outside, and it arrived by luck.
7.3.7 is an attempt to make it not a matter of luck.
The Agent Estate Standard v1.0 is live at aideliverydiscipline.com/standard. It is free, it stays free, and it is licensed so you can implement it and cite it without asking me.
I want to be careful about what it claims, because the field has done serious work this year and this Standard sits on top of that work rather than against it.
Discovery is solved. Microsoft Agent 365, Workday ASOR and ServiceNow will tell you what is running inside your organisation, and they will do it better than any process you could build yourself. Between January and February, four significant agentic governance frameworks published within six weeks of each other. Add ISO 42001 already in force, the EU AI Act, and CPS 230 for the regulated entities among you. Between them they answer whether an agent may operate: whether it is safe, lawful, secure and inside the controls you have accepted. Those are the right questions and I am not going to answer them again.
The question none of them answers is whether an agent is still worth operating, and who answers for it by name.
The registry tells you what is running. The Standard tells you what is worth running.
Conformance has three levels. Level 1, Visible: do you know what is running? Level 2, Accountable: does a named person answer for each one? Level 3, Stewarded: does the estate operate on a cadence, with evidence?
Every requirement across those three levels can be satisfied by producing a document. Except one.
You can write a promotion gate. You can define a review trigger with a numeric threshold against each of the five control surfaces. You can run an estate review every quarter, minute it, and file the minutes where an assessor will find them. Six months of that produces a folder which looks like conformance from any reasonable distance.
You cannot write having stopped something.
The failure signature is in the document itself. Reviews take place and nothing is ever stopped. An estate review that has never retired an agent is a status meeting.
Gartner predicted in June 2025 that more than forty per cent of agentic AI projects would be cancelled by the end of 2027, on escalating costs, unclear business value and inadequate risk controls. The number gets quoted constantly and almost always without its date, which is a pity, because the date is the interesting part. We are past the halfway mark of that window.
Cancellation is not retirement, though. Cancelling a project that never reached production is a budget decision, and organisations are good at budget decisions. Retiring an agent that is running, that people have built their week around, that has somebody's business case attached to it, is a different act. Almost nobody has practice at it.
The strongest argument against 7.3.7 is that it punishes good behaviour. An organisation with a small, young, well-run estate could have named owners, numeric triggers, a real review cadence, and nothing that deserves to be stopped. Should it fail Level 3 for the absence of a retirement it had no reason to perform?
I have thought about this more than any other line in the document, and I have left it as it is, for three reasons.
The requirement says retired or demoted. Demotion returns an agent to supervised or assistive autonomy. Any estate running twelve months with genuine numeric thresholds across five surfaces will demote something, because thresholds set honestly get breached.
An estate that young cannot claim Level 3 anyway. Requirement 5.2.4 gives an adopting organisation twelve months to bring existing agents to promotion-record standard, and 7.3.7 looks back twelve months. The timing does that work already.
And sitting at Level 2 is not a failure. It is the level almost nobody currently holds. An organisation with everything else in place has not fallen short of Stewarded. It has not yet accumulated the track record that Stewarded describes.
There is a version of this requirement that tests action only when a threshold is breached. I considered it and rejected it, because the thresholds are set by the organisation being assessed. Set them loosely, breach nothing, and the requirement is satisfied by having nothing to satisfy. That is the failure mode 7.3.7 exists to close.
Every other requirement in this Standard can be met with a document, which is cheap. This is the only one where the cheapest way to pass is to do the thing.
Here is the part I did not expect when I wrote the conformance section.
The barrier to 7.3.7 has very little to do with retirement.
The levels are cumulative, so a Level 3 claim requires Level 2 to hold first. And Level 2 contains a requirement that reads as though it should be trivial. The count of agents in production without a Business Owner must be zero.
Field 9 of the Agent Record holds the Business Owner. The Standard states that a team, a role, a committee or a distribution list must not be recorded in that field. It has to resolve to a person.
That one constraint is what turns a populated register into an empty one.
There is a number worth pulling out of your register this week. Call it the orphan count. The number of agents running in production today with no named individual accountable for the outcome they were funded to produce.
At Level 2, that number has to be zero.
A team recorded as owner sounds like coverage and behaves like nobody, and not because the team is unwilling. When the estate review asks whether this agent should continue, a team cannot answer. A committee cannot be convened inside the ten working days the Standard allows after a trigger breach. A distribution list cannot be held to a decision six months later.
Retirement is a decision with authority behind it, and authority sits with a person. An estate review that has never retired anything usually has nobody in the room who could.
So the two failures are one failure at different heights. The register whose owner column reads "Data Platform Team" and the review that has run for two years without stopping anything are the same condition, and the first one causes the second.
It is also why the telco needed a new CFO. Not because nobody in that business was capable of asking the question, but because asking it was nobody's job, and a question that belongs to nobody eventually gets asked by an accident.
Requirement 5.5.1 says retirement criteria must be defined at promotion, not at first review. I believe that is right, because criteria written afterwards get written to fit whatever already happened.
I also know it asks something difficult. At promotion, nobody knows enough to write good retirement criteria. You are being asked to describe the conditions under which a thing you have not yet run should be switched off.
The Standard allows the criteria to be revised at review, provided the reason is recorded. That is a compromise, and I would rather name it than pretend it is elegant. If you find a better formulation once you have tried it in practice, I want to hear it.
Appendix A of the Standard sets out the Five-Agent Test properly and it takes about a week. Here is the twenty-minute version.
Open whatever register you have. Take the five agents that matter most by cost, volume or exposure, and look at one column only. Who is accountable for the outcome that agent was funded to produce.
Count how many name a person. That is your orphan count on a sample of five, and it is the most honest number you will see this quarter.
If you have no register at all, the answer arrives faster. Ask three people which agents are running in their business unit and compare the lists.
If your organisation has never retired an agent, what would have to change before it could?
Not whether it should. What would have to be true first. The name that would need to sit in field 9, the forum that would need the authority, the person who would have to agree to lose something.
One or two sentences is enough. If you have retired one, tell me what forced the decision instead, because that is the thing I have never once been told.
I read every reply myself.
vijayan@aideliverydiscipline.com
The Agent Estate Standard v1.0 is at aideliverydiscipline.com/standard. The Agent Record template and the Five-Agent Test are there with it. Free, permanently.
New issues arrive by email first. Subscribe here.
