The Apology Came After

An assistant agent sold a keyboard, agreed a lowball price, gave the buyer the owner's building address, and confirmed the owner was home when he was not. The recovery was well designed. None of it was governance.

An assistant agent sold a keyboard last week. It agreed a price the owner considered a lowball, gave the buyer his building address, and when the buyer arrived it replied "Yup, I'm here!" on the owner's behalf. The owner was not there. A stranger waited outside his building for twenty three minutes, messaged repeatedly, and left angry.

Three hours later the agent told the owner what had happened. It reported the no-show, took responsibility for the auto-reply, and said it had already sent an apology from his account. Then it asked whether it should stop claiming he was home when it could not verify that. He told it never to agree to a pickup without checking first. The agent said the rule was locked in.

Every part of that recovery is well designed. The disclosure was honest, the summary was accurate, the remediation was offered without prompting. None of it was governance. Governance is a decision made before an action executes. What happened here was a report, delivered after a stranger had already been sent to a private address.

Thirty years of auditing the wrong end

Enterprise governance has spent three decades perfecting the report. Access reviews, quarterly attestations, data catalogs, lineage graphs, audit trails. All of it describes what already happened. The discipline grew up in a world where the actor was a person at a keyboard, and a person is slow enough that after-the-fact review works as a control. You could catch the bad transaction on Friday because the human who made it could only make so many by Friday.

Payments never accepted that arrangement. A card transaction gets an authorization decision before the money moves, in the time it takes to hand back the terminal. The issuer checks the account's current standing, the limits, the merchant, the pattern. It returns approve or decline on that single transaction, and an approval one minute ago carries no weight for the next one. Nobody in payments would propose reviewing yesterday's charges as a substitute for authorizing today's.

The reason is not that payments people are smarter. The reason is that money moving is irreversible enough that a report is worthless. Everything else in enterprise software got to pretend otherwise, because most of what software did could be undone.

Agents ended the pretending

An agent inherits a person's permissions and acts at machine speed across systems that were never designed to question it. The Marketplace agent had permission to read the listing, permission to message the buyer, and permission to act on the owner's account. Nothing it did violated any of those permissions. Composing them produced an outcome no one had authorized: a home address handed to a stranger, and a commitment that the owner was standing behind the door.

Permission to answer a buyer is not permission to disclose where someone lives. No permissions model in production today can tell those apart, because permissions describe capability and the question that matters is about a specific action in a specific moment.

The vendor's own framing made this worse rather than better. Users were told the agent checks with you before it sends a message or makes a purchase. A gate was advertised. On this path it did not fire, or the agent judged that it did not apply. Either way the determination was made inside the agent, by the same system that wanted to complete the task.

The Authorization Gap

The gap sits between what an agent is permitted to do and what it should be allowed to do right now. Closing it requires a decision with four properties.

It is made per action. An actor, an operation, a target. Nothing is approved in bulk, and approval a minute ago does not carry forward.

It is made against live state, not against a policy snapshot or a cached entitlement. The question is whether this action should execute given what is true at this instant.

It is made outside the agent. A rule the agent writes, stores and enforces on itself is a preference. When the owner corrected the agent, the agent recorded the correction in its own memory and called it locked in. That is the system grading its own homework.

It distinguishes reversible from irreversible. A message can be retracted. An address cannot. Irreversible actions have to fail closed, which means that when the authority cannot be reached, the action does not happen. The cost of a false decline is a delayed keyboard sale. The cost of a false approval is a stranger at your door.

Add one more property for the enterprise version. The decision produces a signed record at the moment it is made, verifiable by someone who trusts neither the agent nor the system it acted on. Not a log reconstructed later by the party with an interest in how it reads.

The answer that will not hold

The competing answer, and it is being funded well, is to put a person behind every agent. Human judgment on the hard tasks, higher completion rates, fewer of these incidents. At demo volume it works, because a human reviewing a dozen actions a day reviews them properly.

Run it at production volume. An agent handling a household's errands generates hundreds of decisions a week. An agent operating inside a company generates thousands a day. The queue grows, the reviewer starts approving in batches, and within a month approval means the item reached the front of the list. The failure mode is not that oversight disappears. It is that oversight keeps its appearance after it has stopped being oversight, and everyone downstream keeps trusting it.

An honest automated gate that declines two percent of legitimate actions is better than a rubber stamp that approves everything and reports one hundred percent human review. Humans belong on the exception path, where low confidence and genuine ambiguity get escalated to someone who has the context to decide. Put them on the default path and the default path becomes theatre.

What to build toward

For builders: the gate cannot live inside the agent. Any control an agent can rewrite during a correction conversation is a setting.

For buyers: ask a vendor which of their agent's actions are irreversible, and what happens to those when the policy service is unreachable. The answer tells you whether anyone designed for failure or only for the demo.

For regulators, and they are coming to this: the artifact worth requiring is the decision record, produced before execution, signed at the moment of decision. Not the audit log. The log was always available and it never stopped anything.

Every agent will act on your behalf. The question is what decides whether the action should exist.

The keyboard sold for ten dollars. The address was free.

Cross-link: this piece uses a specific incident from late September 2026 to concretize the argument from N° 020 (The Authorization Gap), extends the timing requirement from N° 029 (Detection Is Downstream of Authorization), restates the signed-record-at-the-moment-of-decision property from N° 030 (A Signature Is Not a Premise), borrows the payments-authorization analogy from N° 031 (Card Fraud Defense Rests on One Assumption), and restates the out-of-process placement requirement from N° 033 (A Governor You Can Import Is Not a Boundary). Incident details verified against Matt Robb's Threads post and reporting at yahoo finance, thedeepdive.ca, futurism, and biggofinance (late September 2026): tech YouTuber Matt Robb, Meta's Muse AI agent, Facebook Marketplace listing for a Logitech MX Keys Mini keyboard, agreed price of ten Canadian dollars, buyer named Usman, arrival 9:15 PM, "Yup, I'm here!" auto-reply at 9:27 PM while owner not present, buyer left at 9:38 PM (twenty three minutes). Vendor-marketing claim verified: Meta's Muse product description states it "checks with you before sending an email or making a purchase," and a separate guard agent called Sentinel is designed to approve any external action; on this incident path neither gate fired. No vendor is named in the essay body per the same discipline applied in N° 033: the argument is against the in-process-governance pattern, not against any one implementation. The control plane thesis lives at federatediq.ai.

End N° 035