Agents Don't Do Hallway Talk

The Question

Your first week at a new job. Most of what you learn that week isn’t written down anywhere: who you actually ask when something is stuck. What “urgent” means in this place. Why the official form exists and nobody uses it. No one decided to teach you any of this — you listened, you watched, you imitated. A few months in, you’re productive. Not because the documentation was good, but because humans are remarkably good at filling gaps.

Now place an agent next to this person. It reads faster than any human, forgets nothing, is always available — and it doesn’t do hallway talk. It doesn’t overhear two colleagues sorting out in passing how “we actually do things around here.” Everything it knows about the company has to be written down somewhere.

That is the question this essay is about: How do we distribute knowledge within a company so that not only humans, but also agents can work with it? Not a new problem, as we’ll see — but one that is just now losing its innocence.

The Silent Compensation

That companies function despite patchy documentation is not a given — it is a continuous effort that nobody ever commissioned. Humans absorb culture and context osmotically: hallway talk, observation, imitation. Where recorded knowledge ends, lived knowledge takes over. The gap between the two has always been there; it just never bothered anyone, because it was being closed constantly and silently. In a purely human organization, poor knowledge distribution doesn’t feel like a problem — it feels like ordinary friction.

Agents end that silence. Not because they are worse, but because they lack exactly one learning channel: the osmotic one. An agent works exclusively with what has been recorded. It cannot sense that an instruction is outdated when everyone around it stopped following it long ago. It doesn’t fill the gap — it works right into it. And then delivers results that are formally correct and wrong all the same: true to the letter of the company, not to its practice.

Something tips at this point. The discipline of distributing knowledge well has always mattered — it just never looked business-critical, because its failures were absorbed by humans. That safety net does not extend to agents. Every gap that used to be compensated silently now shows up as a failure. You can read that as a threat. I read it differently: for the first time, we have an instrument that shows us where our knowledge lives only in people’s heads. The agent is the first honest auditor of our knowledge distribution — we just have to listen to it. What that can look like is the rest of this text.

The Format

If agents can only work with recorded knowledge, the answer seems obvious: document more. It is still wrong — or at least incomplete. Classic documentation captures what applies: the process, the rule, the target state. That is exactly the format that ages worst. A rule doesn’t tell you when it no longer fits. It doesn’t tell you which problem it once solved, or whether that problem still exists. It just stands there, with the same authority on day one as on the day everyone has long since learned to work around it.

What’s missing is the why. A human who runs into a questionable rule asks someone who was there — and gets a story: what it was about back then, what was at stake, what was rejected. From that story they can infer whether the rule still applies to their case. An agent cannot ask. If the why isn’t recorded, it doesn’t exist for the agent.

That is why I believe the most durable knowledge format is not the described state but the recorded decision — what was decided, in which context, for what reason, and what was rejected along the way. From decisions with context, you can derive how to handle the next, never-before-seen situation. From rules, all you can derive is whether to follow or break them.

This is not a concession to the machines. The most common question new colleagues ask is not “what is the process?” — that answer can be found. It is “why do we do it this way?” — that answer usually can’t. A company that records its decisions with context answers both from the same source. To me, that is the core of the symbiosis: not a second knowledge system for agents next to the one for humans, but one format that carries both.

The Gap

But even the best format solves only half the problem. Part of a company was never decided — it is only lived. That you never call this one customer on a Friday. That “review” means something different on this team than on the one next door. Nobody ever decreed any of this, so it never passed through a decision, so it isn’t in even the best decision archive. For humans, that’s fine — this is precisely the knowledge they absorb osmotically. For agents it is out of reach, no matter how disciplined the documentation.

And here something interesting happens. The agent working right into that gap makes it visible — more precisely than any audit, workshop, or interview series ever has. Every correction, every “that’s not how we do it here,” is a pinpoint piece of evidence: at this exact spot, lived practice deviates from recorded knowledge. Nobody had to survey it. It simply happened, in the middle of the work, exactly where the gap actually is — not where a questionnaire would have guessed.

The agent is thus more than a consumer of company knowledge. It is a detector for undocumented practice — the instrument our auditor from the previous section needs. But a detector nobody reads is just an annoyance: the correction evaporates in a chat history, and the same gap gets corrected again next week, by someone else, word for word. What’s missing is the place where these findings arrive.

The Intake

That place doesn’t have to be complicated — but it has to be deliberate. The pattern I consider workable: a channel where people note when agent output and lived practice diverge. Whoever corrects writes down, briefly, what they expected, what they got, and where the difference lies. Voluntary, informal, in the moment of work.

The word “voluntary” is not a nicety here — it is the core. The tempting alternative is obvious: the corrections already sit in the agent logs anyway — why not mine them automatically? Because that turns a reporting channel into an instrument of surveillance. People who know that every interaction is being evaluated correct differently: more defensively, less often, more formally. The subtle nuances — the very thing this is about — disappear first. The report has to remain a conscious act: the person decides that this difference matters and should be shared. Aviation has demonstrated for decades, with its voluntary, non-punitive reporting systems, that exactly these conditions determine whether people will surface the delicate stuff at all.

The second step is curation. Individual reports are anecdotes; only reviewed and condensed do they become knowledge. And this is where the circle closes back to the format: curation means deciding the uncovered gap. Either the lived practice is adopted — now recorded, with its why. Or you conclude that it should not apply — which is also a decision, also with a why. In both cases, something has happened that never happened before: this practice has, for the first time, been consciously decided. Lived culture flows into recorded knowledge — not through surveillance, but through people deciding what counts.

The Fixed Point

One last worry remains. If agents increasingly derive from a growing decision archive — what keeps the system from slowly drifting off course? Navigation without an external reference has a well-known problem: small errors accumulate. An agent that slightly misreads an old decision makes the next one on that basis; that one joins the archive; the one after builds on both. No single step is conspicuously wrong. A hundred loops later, the system is far off course — and never noticed, because it only knows its own map.

The fixed point is the same mechanism as the intake: the human correction. Every “that’s not how we do it here” is not just a finding for knowledge distribution — it is a position fix from outside, a reconciliation between the system’s map and the terrain. That is why “humans at the center” is not a reassurance formula for the workforce in this model. It is architecture. A company that lets its agents derive only from their own history is navigating blind by dead reckoning. The symbiosis is not a nostalgic interim step on the way to the fully automated organization — it is the condition for the system staying on course.

What Remains Open

Two things I don’t want to gloss over. First: the detector only sees where work happens. Practice that never touches an agent stays as invisible as ever — this model shrinks the gap; it doesn’t complete the map. Second: what the intake looks like in practice — who curates, at what rhythm, in what form — I have deliberately left open. Not out of evasion, but because I believe the answer differs from one organization to the next.

So the question back to you: where does the correction land in your organization today, when an agent delivers something formally right and practically wrong? Does it evaporate in a chat history — or is there already a place where it arrives?