There is no written artifact I can ask for anymore that proves the person responsible actually understands the requirement. Tickets, specifications, pull requests, a long explanation in Slack: all of them can be coherent, well organized, and correct on the surface while nobody involved holds the mental model behind them. Once a model can generate every asynchronous artifact, the artifact no longer proves anything about the human attached to it.

This goes for me, too.

What the Written Artifact Used to Tell You

A ticket broken down into clean phases used to tell you something about the person who broke it down. Not because the writing was good, but because producing it required carrying the problem around long enough to see its shape. The document was a side effect of understanding, so you could read the document and infer the understanding.

That inference is what broke. Ask someone to take a high-level task and decompose it, and what comes back may be a real decomposition or it may be generated content that happens to read like one. Then you are sifting through generated tickets and generated PRs trying to recover intent from output. Anything async has this problem, the same way an interview does: you are reading an artifact and guessing at the person behind it.

This Is a Management Problem

AI does not make a misunderstanding smaller. It builds faster and further on top of it, so a misunderstanding nobody caught early ships before anyone gets around to asking about it. That is what raises the stakes on whether the human in the middle actually understands the thing.

If someone sits between a leader and a set of agents, that position has to earn itself. It earns itself by questioning the assumptions in the request, surfacing the architectural forks, naming the risks, and owning a recommendation. Those are the things that are hardest to hand off, because they require holding the requirement and the system in your head at the same time and noticing where they disagree.

When the person in the middle is relaying generated artifacts instead, that layer is lossy. It adds delay without adding judgment, and at that point it is fair to ask what going through the person is buying that routing the work straight to agents would not. That is an uncomfortable question to put to a role, and I think it is the right one.

Separate Preparation From Verification

The practical response is to split those two things apart rather than to restrict the tools.

Preparation is where AI belongs. Let the owner of a decision use it to challenge their own proposal before anyone else sees it, draft the ADR that records the decision and the alternatives, and build a prototype if the question needs one.

Verification is the part that has to be unassisted: a live review where the owner explains the decision and its tradeoffs without consulting AI during the conversation. That is not a ban on tools, and it is not argument theater. Everyone would be there to find the best decision rather than win, and what such a meeting establishes is narrow: that the accountable human understands the choice well enough to direct agents against it and to recognize when their output starts drifting from it.

Once the ADR is accepted, agents can start implementing it immediately, which makes a two-day rhythm possible: design and prototype on day one, live defense and a decision, then build and merge on day two. Generation is not the scarce resource in that loop, so the thing worth scheduling deliberately is the understanding, the questioning, and the judgment. The generation follows from that.

The Counterargument

The strongest counterargument is that oral fluency is not the same as understanding. Someone can talk well about a decision they do not hold, and live review disadvantages people who genuinely think better in writing than out loud. I do not have a clean answer to that, and I do not think the answer is to trust written records less carefully or to replace them.

Keep the ADR as the durable source of truth. It is still the record you go back to later, and it is still what the agents build against. Use the live questioning as a falsification step instead, aimed at what a prepared explanation does not cover on its own: can the owner name the alternatives they rejected and why, answer an edge case nobody prepared them for, and say what evidence would change their mind?

None of that proves understanding on its own. It is the artifact plus the failure to falsify it, not either one alone.