Agents don't fail because the model is weak. They fail because the context is thin. Wittgenstein figured out why in 1953 — and the AI industry is still making the mistake he diagnosed.
The chain nobody's making explicit
Here's the chain:
Wittgenstein's central insight — meaning is use, not representation — implies that context determines meaning. Context-as-meaning implies that agentic systems fail on thin context, not on thin capability. Building thick context is therefore not an engineering problem. It's a philosophical competency. And that competency is what separates agents that work from agents that produce confident-sounding outputs built on nothing.
Every link in that chain matters. Skip one and the whole argument collapses into either "just write better prompts" or "AI isn't ready yet." Both are wrong. The technology works. The meaning architecture around the technology doesn't.
What Wittgenstein actually said
Ludwig Wittgenstein spent the first half of his career building a theory of language as logical representation — words as pictures of facts, propositions as mirrors of reality. The Tractatus Logico-Philosophicus. It was elegant, rigorous, and wrong.
He spent the second half dismantling it. In Philosophical Investigations, published posthumously in 1953, he replaced the picture theory with something radically different: meaning is not the relationship between a word and the thing it points to. Meaning is how the word gets used in practice. "The meaning of a word is its use in the language."
This sounds simple. It isn't. It dissolves a mistake that most of the AI industry is still making.
The representationalist view — meaning lives in the symbol itself — is the default assumption behind prompt engineering. Choose the right words, arrange them correctly, and the model will understand. This is the picture theory applied to language models. It doesn't work, for the exact reason Wittgenstein said it wouldn't: meaning lives in use, not in the symbol.
His most famous thought experiment makes this concrete. A builder and an assistant work together. The builder calls out "slab!" and the assistant brings a slab. "Slab" doesn't mean slab because it represents the concept of a slab. It means slab because of the shared practice — the work being done, the roles being played, the established pattern of use. Change the practice and "slab" means something different. Move the language game to a philosophy lecture and "slab" becomes a reference to Wittgenstein.
Same word. Different meaning. The situation is doing all the work.
The insight isn't unique to Wittgenstein — Fillmore's frame semantics and Clark's work on common ground arrived at the same conclusion from linguistics and pragmatics respectively. Meaning doesn't live in symbols. It lives in the situated, shared context that gives symbols their function. And that insight is exactly the one the AI industry needs and doesn't have.
The context engineering gap
Strip away the hype and this is what "context engineering" actually describes: providing enough situational information that the language model can participate in the right language game.
I wrote in January that context engineering isn't a discipline — it's a literacy. I stand by that. But there's a deeper claim underneath it that I didn't make explicit: the reason context engineering matters is philosophical, not technical. The model fails on the situation, not the syntax. And situations can't be represented — they can only be described, with all the specificity and structure that description requires.
This is why dumping an entire codebase into a context window produces worse results than providing the relevant file with a clear description of what you're trying to do. More information isn't more context. Information becomes context only when it's structured to convey the situation — the practice, the intent, the constraints, the relationships between the parts.
The AI Operating System piece I published last week identified the gap: 94% AI capability meeting 33% deployment. Most businesses can't describe what they do in terms precise enough for any system — human or machine — to execute reliably. That's not a technology failure. It's a clarity failure. And clarity about what your business does, how decisions get made, what matters and what doesn't — that's a question about meaning, not about engineering.
The businesses getting the most out of AI aren't the ones with the best tools. They're the ones that have already done the philosophical work of making their practices explicit. They just don't call it philosophy. They call it process documentation. But the underlying operation is identical to what Wittgenstein described: translating implicit know-how — the kind of understanding you have when you know how to go on — into explicit context that someone, or something, else can use.
Meaning architecture
This needs a name. I'm calling it meaning architecture.
What is meaning architecture? The deliberate structuring of context — organizational knowledge, decision criteria, process logic, relational information — so that an AI system can participate in the correct language game. Not the right prompt, not the right model: the right situation, encoded in a form that makes the intended meaning operational.
Meaning architecture is the deliberate structuring of context — organizational knowledge, decision criteria, process logic, relational information — so that an AI system can participate in the correct language game. Not the right prompt. Not the right model. The right situation, reconstructed in a form that makes the intended meaning operational.
A CLAUDE.md file is a small piece of meaning architecture. It tells the model what project this is, what conventions apply, what the directory structure means. But the real work is larger. It's the knowledge graphs, the process documentation, the decision frameworks, the institutional memory that together constitute what a business means when it says "handle this customer inquiry" or "evaluate this investment opportunity" or "build this feature."
When those structures are thin — a few bullet points, some vague instructions, an incomplete process map — the agent generates outputs that sound plausible and miss the point entirely. The model isn't broken. The meaning architecture is absent. The agent is playing the wrong language game because nobody specified which game it should be playing.
When those structures are thick — comprehensive context about the business domain, explicit decision criteria, clear descriptions of roles and relationships, institutional memory that captures not just what happened but why — the agent produces outputs that feel like they understand the situation. Because, functionally, they do. Not in some deep philosophical sense. In the Wittgensteinian sense: they know how to go on.
The trust contraction
There's an economic implication here that goes beyond tooling.
When AI commoditizes capability — when anyone can spin up an agent that writes, codes, analyzes, and strategizes — the bottleneck shifts. It's no longer "who can produce the best output." It's "who do I trust to understand my situation well enough to point the AI in the right direction." That's not a capability question. It's a relationship question. And relationships don't scale past Dunbar's limit.
The scarce resource isn't intelligence. It's trust. The world contracts back toward small, high-context networks — not because technology limits us, but because trust doesn't scale. It never did.
Epistemic differentiation
This has a direct implication for anyone building a business in the agentic infrastructure space.
Most consultancies differentiate on capability. "We're better at X." But capability differentiation is collapsing. If AI does X, being better at X is a temporary moat that erodes with every model release.
The next layer is methodology. "We have a proprietary process." Better than capability, but still replicable. Reverse-engineer the process, encode it, ship it as a product.
What survives is epistemic differentiation — a different way of understanding the problem itself. Not "we build better agents" or "we have a better process." Rather: "we understand that agentic systems are context systems, and context is a problem of meaning, not a problem of engineering."
That's the gap. The entire agentic infrastructure industry treats context as a data problem — how to get the right information into the right window at the right time. RAG pipelines, vector databases, knowledge graphs, memory systems. The tooling is real and the engineering is impressive. But it addresses the wrong level of the problem. Getting information into the context window is plumbing. Knowing which information constitutes the meaning of the situation — that's architecture.
The distinction tracks Wittgenstein's critique exactly. The representationalist assumes meaning is in the data. The pragmatist knows meaning is in the use. Building a system that retrieves relevant documents is engineering. Building a system that understands what "relevant" means for this business, this client, this decision — that's meaning architecture.
And meaning architecture can't be commoditized, because it requires the one thing AI can't replicate: deep contextual understanding of a specific situation, earned through relationship and accumulated over time. The model layer is indefensible. The meaning layer — the layer that tells the model what this business is, what matters, what the words actually mean in practice — is the moat.
What the market selected for
The Anthropic-Pentagon standoff accidentally tested this thesis. The market was supposed to punish the model with boundaries and reward the compliant one. The opposite happened — churn dropped from 55% to 36%, enterprise adoption nearly doubled, and ChatGPT uninstalls spiked 295%.
The market didn't select for sycophancy. It selected for trust. A system that evaluates its own outputs against explicit principles — that has a defined relationship to what it generates — is a system people trust with consequential work. That evaluation layer is meaning architecture applied at the model level. And the same design principle runs through every trustworthy infrastructure, from proof-of-work to constitutional AI.
People don't want a dog. They want a colleague with their own framework for evaluating quality. That requires more meaning architecture, not less.
What structures form
The trust contraction reshapes organizations around this logic. The firm shrinks to the minimum viable trust network. Everything that used to require headcount for capability reasons becomes infrastructure. What remains is the irreducibly human part — the contextual understanding that tells the infrastructure what the situation actually means.
The people who build that meaning architecture are the bottleneck. Not because they're smarter than the machines. Because they understand a specific situation in a way that can't be downloaded, distilled, or automated. It was earned through relationship, built over time, and maintained through trust.
That's not a comforting story about how humans will always matter. It's a structural claim about where value concentrates when capability is commoditized.
What remains
The skill isn't "knowing things" or "producing things." It's thickening threads — taking a thin conversation and making it thick. Asking the question that unlocks the next level of context. Naming the thing that's been circling unnamed. That's a philosophical competency: creating vocabulary, dissolving confusion, reframing questions until the shared context is dense enough for agents to operate on.
Meaning architecture is the skill. Trust networks are the structure. Thick threads are the practice.
The people who build meaning architecture for others don't just architect it and hand it over. They teach others how to thicken their own threads — how to build the common ground, articulate the frames, and develop the situated vocabulary that makes agents work. Not a deliverable. An ongoing practice. A language game that gets richer every time you play it.
Sources
- Wittgenstein, Ludwig — Philosophical Investigations (1953) — §43: "The meaning of a word is its use in the language."
- Wittgenstein, Ludwig — Tractatus Logico-Philosophicus (1921)
- Fillmore, Charles J. — "Frame Semantics" — Linguistics in the Morning Calm — 1982
- Clark, Herbert H. — Using Language (1996) — Common ground and collaborative communication
- Anthropic — "The Impact of AI on Labor Markets" — March 5, 2026
- Apptopia — "Gen AI Chatbots: February 2026 Data Brief" — Claude churn rate 55% → 36% (August 2025 to February 2026)
- Sensor Tower — "ChatGPT Uninstalls Surge Amidst Deal With US Department of War" — U.S. app uninstalls +295%, February 28, 2026