The knowledge nobody wrote down
Every company has a system that seven people know how to use. That knowledge is the reason most automation stalls, and it is the thing an agent has to be given before it can be trusted with anything.
Somewhere in your company there is a platform that exactly seven people know how to operate. Not because it is complicated. Because the part that matters was never written down.
Which report is stale by Thursday. Which button silently fails and has to be clicked twice. Which two steps have to happen in that order or the whole thing rolls back. What the acronym on the third tab actually stands for, and why the team that named it has since left.
None of that is in the documentation. It is in people.
Why this breaks automation, over and over
This gap has quietly ended every generation of enterprise automation. Scripted workflows and rules engines both tried to route around it, and both hit the same wall, which a research team at Distyl put in one line: you cannot automate what you cannot articulate. (Environment Maps, Distyl, June 2026.)
AI agents hit the same wall, and it is worth being precise about why, because the failure looks like stupidity and is not.
An agent can see the interface perfectly well. What it lacks is a persistent model of how the environment behaves. So it reconstructs its understanding from scratch on every single task, from whatever happens to be in the prompt that time. It has no memory of the button that fails, no memory of the ordering constraint, no memory that this worked last Tuesday. It is a competent operator with amnesia, meeting your systems for the first time, every time.
That is why agents perform well on demos and badly on real work. A demo has a prescribed path. Your business does not.
What actually fixes it
The Distyl team's answer is what they call an Environment Map: a persistent knowledge graph that acts as a navigational aid for the agent. Not more text in the prompt. A structured model that survives the task.
It holds four things:
- Contexts. Where the agent is.
- Actions. What it can do from here.
- Workflows. How a task actually gets done, in what order.
- Tacit knowledge. What the experts know and never wrote down.
Their reported result: agents equipped with one achieved close to double the task success rate of agents without, in complex web environments. That is their measurement, from a paper presented at the ICLR 2026 workshop on world models, not ours.
The mechanism is the interesting part, and it generalises well beyond web agents. Most systems handle context by stuffing more information into the prompt, where it evaporates the moment the interaction ends. Structure it instead, and it persists, accumulates, and compounds. The agent starts following known paths rather than rediscovering them. It gets more predictable, not less, the longer it runs.
Why we are writing about someone else's paper
Because it is the same argument Voltaire is built on, arrived at from a different direction, and independent confirmation is worth more than another sentence from us.
We say decisions come out of a knowledge graph of your business and a formal-logic engine, not out of a language model. The reason is not architectural taste. It is that a language model asked the same question twice can answer twice, differently, with no way to audit either answer, because nothing persisted between the two. Structure is what makes an answer reproducible, and reproducibility is what makes it auditable, and auditable is the minimum bar for software that is allowed to move your budget.
Distyl reached the same place by watching agents fail at enterprise tasks. We reached it by asking what it would take to stand behind a recommendation six months after making it.
What this means if you are buying
There is a question worth asking any vendor whose software is going to touch your operations, and it separates the two kinds of system quickly:
What does your system remember about my business between one task and the next, and where does it live?
If the answer is a longer prompt, a bigger context window, or a document that gets re-read each time, you are buying an operator with amnesia. It will be impressive in the demo and disappointing in month three, and the reason will be structural rather than a matter of tuning.
If the answer names a persistent structure, ask the follow-up: what happens when your interface changes? A model built to accumulate adapts. A prompt has to be rewritten by whoever remembers what it said.
The uncomfortable part, for buyers and vendors alike, is that building that structure means someone has to sit with the seven people who know how the platform works and get it out of their heads. There is no version of this where the tacit knowledge extracts itself. Any vendor promising otherwise has not done it yet.