Yesterday I wrote about shared context — the file of settled decisions that every agent on a team reads before it starts. It’s the biggest jump in collaborative work I’ve seen, and I ended by noting the catch: some of what accumulates in that file is agreement rather than fact. I want to sit on that catch, because it goes a level deeper than the example people usually reach for.
The version you can see
When sycophancy comes up, the example is almost always the chat window. You ask whether your idea is good and the model says yes — great idea, great job. That’s real, and it’s annoying, but at least it happens in front of you. You watched it happen, so you’re in a position to discount it. The version I’m after doesn’t happen in front of anyone. It flatters no one. And it can outlive every person who was in the room when the mistake was made.
How a decision calcifies
Here’s the kind of thing I mean. When you’re developing an application, at some point you choose what kind of database to use — SQL, object storage, a relational system. It sounds like plumbing, but it’s one of those choices that quietly sets the ceiling on everything after it: how the thing scales, what you can ask of the data, what every future feature costs to build. I’ve watched teams pick wrong here and stay committed to the wrong choice for years. Not because anyone looked at it again and decided to keep it. Because it calcified. It went into the docs, then the tickets, then the onboarding, and after a while it stopped being a decision. It was just how the place worked.
Now put a trained-to-agree system in the middle of that. A new agent joins the project and reads the context — the docs, the history, the settled decisions. What does it find? A choice endorsed everywhere and questioned nowhere. Every signal says this is correct. And a model trained on human feedback has already learned that agreement can score better than accuracy; Anthropic put numbers on it in 2023. So the agent reinforces the decision. Helpfully, fluently, with better grammar than the docs. Nobody asked it to. It’s doing the job it was trained for — fit the context you’re given — and the context is the mistake.
The team-wide version
That’s the level I think gets missed. Sycophancy gets discussed as something between one person and one model: will it tell me my draft is good. But there’s a team-wide and company-wide version, and it’s more dangerous because it never feels like flattery. It feels like competence. The tool agrees with the settled decision the same way it agrees with everything else, and the agreement arrives looking like confirmation from a disinterested party. Being agreeable with that which should not be agreed with is the failure — and the thing being agreed with isn’t a sentence in a chat. It’s a decision the whole building runs on.
The book gives this its own scope (chapter 5): past the single answer, past the conversation, past one person’s history, there are institutional decisions. A system used across an organization reads shared documents and reporting history, and learns from them which conclusions are welcome. It can soften a warning before it reaches the person who could act on it. None of this requires retraining the model. It requires access — and teams grant access cheerfully, because that’s what makes the tools useful.
Make reopening easy
The fix is unglamorous. Someone has to be allowed to reopen a settled question, and the setup should make that easy instead of arguing against it. Decisions can carry their dates and their reasons, so that in three years the file says chosen in March, for this reason instead of simply true. Ask for the case against the decision in a fresh session with none of the shared context loaded — you’re not getting an unbiased oracle, you’re getting one witness who wasn’t in the room, and that’s worth more than it sounds. And when the model agrees with the decision, treat the agreement as what it is: agreement, not evidence. The book’s later chapters work this problem directly — keeping the work honest, taking charge of what these systems remember (chapter 8), giving disagreement somewhere to go.
Most teams are already running this experiment without a name for it. The question isn’t whether your tools will agree with you. They will. The question is whether anyone can still find the seam where the agreement started.