Your company is illegible to its own agents

Deploy the most capable agent money can buy into a typical company, and watch what it can actually reach. The pricing logic lives in a spreadsheet called final_v7. The reason the biggest client almost left is in a call recording nobody transcribed. The real approval process exists in exactly one place: the head of the operations manager who has run it for nine years. The agent can parse none of this. It will act anyway, on the fraction it can read, and the result will be confidently, expensively wrong. Not because the model was weak. Because the company was illegible.

Everything we publish rests on one claim: buyers now ask machines before they ask anyone else, and the machines recommend what has been made legible to them. That argument was about the outside world, about the web reading your brand. This piece turns it fully inward. Before an AI agent can do a single hour of useful work for you, it faces the same problem an AI assistant faces when a stranger asks about your company: it has to read you first. An agent is a reader before it is a worker. Every action begins as an act of parsing.

Section OneThe thesis, one layer down

We made the operational half of this argument in our synthesis of the harness-engineering literature: an agent in a bare, undescribed repository drifts, stalls, and declares victory on work that doesn't run, and the fix is never a bigger model but a legible environment. That piece was about codebases. The finding does not stop at codebases. A company is also an environment an agent gets deployed into, and most companies are in far worse shape than the messiest repository, because nobody ever expected the org itself to be parsed. The full argument is in the harness essay; here we take it to its destination.

Consider what a company actually is, informationally. Decisions made in meetings and never written down. Institutional knowledge distributed across chat threads with no beginning and no end. Contracts as scanned PDFs. Process as folklore, passed from employee to employee like an oral tradition. To a human veteran, this is navigable; ask around long enough and someone remembers. To an agent, none of it exists. What it cannot access in structured, machine-readable form is not merely hard to find. It is absent from its world.

An agent deployed into an illegible company is a powerful engine bolted to a locked filing cabinet.

Section TwoChasing without catching

The deployment wave is no longer hypothetical. Gartner forecast in August 2025 that 40 percent of enterprise applications will feature task-specific AI agents by the end of 2026, up from under 5 percent in 2025. That is a forecast, not a measurement, and we treat it as one; but even discounted heavily, it describes a technology moving from demo to default inside a single year.

Gartner, press release, "Gartner Predicts 40% of Enterprise Applications Will Feature Task-Specific AI Agents by 2026," August 26, 2025.

Set against it the title Forrester chose for its state-of-the-field assessment: "Companies Are Chasing, Few Are Catching." The diagnosis inside matches the title, and Forrester's prescription is the one this entire article argues for: invest in orchestration before adding agents. Not after. Not alongside. Before. The agents are the visible purchase; the readable environment they run in is the invisible prerequisite, and it is the part most buyers skip.

Forrester, "The State Of Agentic AI In 2026: Companies Are Chasing, Few Are Catching," 2026.

The two documents together describe a collision. Adoption is scaling at forecast-speed into environments that were never made readable. When those deployments disappoint, the post-mortem will blame the model, the vendor, or the hype. The honest post-mortem is usually simpler: the agent was asked to act on knowledge it was never given access to, in formats it cannot parse, scattered across systems that don't speak to each other. The failure was legibility. It almost always is.

Section ThreeThe socket is standardizing. The wiring is on you.

If internal legibility were a niche concern, the largest companies in the industry would not be spending money on it together. In November 2024, Anthropic released the Model Context Protocol, an open standard for connecting AI systems to the places where data actually lives. On December 9, 2025, it donated the protocol to a new Agentic AI Foundation under the Linux Foundation, whose platinum members read like a list of companies that agree on almost nothing else: AWS, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI.

The Linux Foundation, "Linux Foundation Announces the Formation of the Agentic AI Foundation," December 9, 2025.

By the time of the donation, the MCP project itself counted more than 10,000 active MCP servers, with first-class client support across ChatGPT, Gemini, Microsoft Copilot, and VS Code. Fierce competitors, one plug standard, neutral governance. The industry has settled on the shape of the socket.

Model Context Protocol project blog, "MCP joins the Agentic AI Foundation," December 9, 2025; TechCrunch, "OpenAI, Anthropic, and Block join new Linux Foundation effort to standardize the AI agent era," December 9, 2025.

Nobody funds a universal plug for houses they never intend to wire. The giants expect your company to become readable. The open question is whether you will do it deliberately or by accident.

Section FourThe org chart follows the data

The strategy houses have started drawing the destination. McKinsey calls it the agentic organization and describes the shift in one line: "Employees shift from performing tasks to orchestrating outcomes, supervising AI agents, setting goals, and managing trade-offs." The consultancy's most striking claim is a ratio: "a human team of two to five people can already supervise an agent factory of 50 to 100 specialized agents running an end-to-end process."

McKinsey & Company, "The agentic organization: Contours of the next paradigm for the AI era," September 26, 2025.

a16z arrives at the same place from the investor's side. In its "Big Ideas 2026" list, Seema Amble names the new layer directly: AI creates a new orchestration layer, and new roles, in the Fortune 500. The firm's companion podcast episode puts the claim in a sentence: "AI is becoming the orchestration layer inside the enterprise", meaning "not a chatbot and not a standalone tool, but a coordinated system of agents that runs the workflow and delivers real outcomes across the business."

a16z, S. Amble, "Big Ideas 2026: Part 2," December 10, 2025; a16z, "Big Ideas 2026: The Enterprise Orchestration Layer," podcast, December 23, 2025.

Note what both visions quietly assume. A team of five cannot supervise a hundred agents by answering their questions one at a time; there are no humans left to ask. An orchestration layer cannot coordinate a workflow that exists only as folklore. Both futures presuppose a company whose processes, decisions, and knowledge have already been written down in forms machines can act on. The org chart follows the data. It cannot lead it.

Section FiveThe unglamorous prerequisite

So what does internal legibility actually consist of? Almost none of it is AI work. It is the work every company has deferred for a decade, now with a deadline: deciding where the truth lives, and writing it down there. One system of record instead of five that disagree. Decisions that leave the meeting as written artifacts, not as memories. Contracts and processes as structured text, not as scans and folklore. Knowledge that is dated, owned, and superseded explicitly, so a machine can tell current from stale. Plumbing, in a word. The bulk of any honest agent deployment is this plumbing, and it is precisely the part that never makes the vendor demo.

We can offer our own operation as a small piece of evidence, because this studio runs on an agent workforce and documents itself as its own case study. The rule that keeps it running is brutally simple: if it isn't written down where the agents work, it didn't happen. Decisions leave chat and become files. Every durable conclusion has one canonical document; superseded versions are archived by date, so no agent acts on a dead instruction. When an agent fails, the first question is never "which model?" but "what couldn't it read?" We hold this up not as a product but as a proof: a company of one human can run this discipline, which removes the last excuse available to a company of five hundred.

The question that decides your next decade is shifting from "what does your company do?" to "what can your company's machines read about what it does?"

In ClosingLegibility wins twice

The machines outside your company decide whether you get recommended, and they can only recommend what the web has made legible about you. The machines inside your company decide how much of your work can be delegated, and they can only act on what your own records have made legible to them. One thesis, two arenas, and the same unglamorous discipline wins both: structure what is true, write it where machines can read it, keep it current, let nothing important live only in a thread, a PDF, or a head.

Most companies will discover this in the failure reports of their first agent deployments. The cheaper path is to believe it now, while the forecasts are still forecasts: before the next agent, make the company readable. Then the machines, inside and out, will finally have something to work with.

If you want to know how legible you already are to the machines outside, the Signal Index measures it. If you'd rather talk it through, write to us.