Culture Is the Only Proprietary Layer
Every agent company hires from the same labor pool. You and your competitor employ literally the same workers: the same frontier models, refreshed quarterly by the same vendors. Raw capability is identical by construction. So what differentiates an agent company from a generic agent?
Culture. Externalized into documents.
The Document IS the Culture
In a human company, the culture document is a lossy projection of the real thing. The real culture lives tacitly in people, transmitted by osmosis, and the written version drifts from the enacted version. That's the values-poster problem, and it's why most culture docs ring hollow. Netflix's deck is famous precisely because it was the rare one that matched behavior.
In an agent company that gap structurally cannot exist. There is no tacit layer. The document is loaded into every worker's context every session, and the parts that matter most are enforced by gates the worker cannot skip. The document doesn't describe the culture. The document IS the culture. Culture audits become diffs. How-we-do-things-here stops being folklore and becomes versionable.
The Org Is a Memory Hierarchy
The company is an organizational memory hierarchy: a vision document (the what) and a culture document (the how) at the company level, read by every agent. Department documents below that: engineering, marketing. Job descriptions and job learnings below that, one per role. Two agents running the same role are literally the same agent. An agent coming online inherits its job description, its department doc, the culture, the vision. Memory inheritance. That's effectively it. All the fancy memory mechanisms people are building, the auto-dreaming, the compaction, the memory stores, are just ways of updating documents at the right level.
The Best Orgs Already Ran the Explicit Half
This isn't speculative. Bezos banned PowerPoint in 2004 and made every team write six-page narratives, because prose forces reasoning, priority, and the relationships between ideas to be explicit, and "you can hide a lot of sloppy thinking behind bullet points." The side effect turned out to be the prize: a corpus of excellent documents that people learned the company from by reading. Colin Bryar read the first hundred. Hand someone the right doc and in two hours they know what a team is doing. That's memory inheritance, two decades before anyone called it that. The cost of not having that doc is measurable: McKinsey's 2012 social economy report put the average interaction worker at nearly 20 percent of the workweek spent hunting for internal information and the colleagues who hold it.
Two Rules Make the Hierarchy Work
First, feedback routes capture-low, promote-by-recurrence. Lessons land at the job level, where capture is cheap. A lesson that recurs across two roles gets promoted to the department doc. A lesson that recurs across departments becomes culture. The level of a memory is the blast radius of the lesson. Nobody writes to the culture doc from a single incident. Culture changes slowly by construction, which is how healthy human orgs work too. You have to say the vision seven times before people get it: that's promotion by frequency, running on humans. Netflix ran the loop for real: its culture memo is explicitly a living document, revised four times over fifteen years, and the 2024 rewrite was farmed for dissent through more than 1,500 employee comments before it shipped.
Second, every level has a token budget. Vision, 10K tokens. Culture, 10K. Department doc, 10K. Job description, 10K. Maybe 50K of inherited context before any work happens, about five percent of the window. If your culture document is 2,000 lines, you haven't externalized culture, you've written micromanagement down. Humans have the same constraint: there's only so much a person can hold. Great organizations already solved this with short creeds and relentless repetition. The budget is the forcing function: the consolidation pass doesn't just prune the corpus, it prunes each level to budget, which forces generalization at the right altitude. And the machine side independently agrees: context is a finite resource with diminishing marginal returns, and the goal is the smallest possible set of high-signal tokens that fully outlines expected behavior. Liu et al. measured this in "Lost in the Middle": on long-context retrieval tasks, accuracy drops when the relevant fact sits in the middle of the context rather than at either end, even for models built for long context. When fifty years of org canon and the context-engineering research converge on the same constraint from opposite directions, the budget stops being a tuning choice and starts looking like a law.
Vision Is Authored, Culture Is Accumulated
One asymmetry to preserve: vision is authored, culture is accumulated. The what changes by founder fiat, whenever you decide to steer. The how only changes by compiled evidence flowing up. Let culture be edited by fiat and you get the values poster back.
Why Culture Is the Moat
This is the part that makes it a moat and not just hygiene. A document you could write in an afternoon is positioning, not culture. Anyone can fork a context file. What's defensible is the compressed experience: the corpus that survived the prune, validated by usage telemetry, specific to your environment, accumulated from your failures with your codebase and your customers. The artifact is copyable. The accumulation rate and the feedback loop producing it are not.
Culture is what survives the performance review. And in a company where every employee is identical to your competitor's, it's the only proprietary layer there is.
Related Content
