Engineering
Inference Logic
Cognitive Agents
In April 2026, a number ignited the AI community: 98%. According to independent researchers analyzing leaked source code from Anthropic's Claude Code project — roughly half a million lines — only 2% of the system constituted what most people would recognize as "AI." The rest was orchestration, scaffolding, data pipelines, error recovery, compliance logic, and interface plumbing. It was, in Matteo Grassi's striking formulation, not an artificial mind but an artificial infrastructure. And yet — this was, in a profound sense, the most honest portrait of intelligence anyone had ever leaked.
1 The Leak and What It Showed
The researchers who analyzed the Claude Code source — Matteo Grassi, Shivam Shrivastava, and Yudistira Ashadi, writing independently between April 3 and April 7, 2026 — were not trying to diminish Anthropic's achievement. They were trying to correct a pervasive misunderstanding: that an AI system is, essentially, a large neural network that receives a prompt and emits an answer.
What Shrivastava described instead was a multi-agent architecture — four cooperating agents, each with a discrete cognitive role:
What strikes a cognitive scientist — or anyone who has read Minsky carefully — is not how novel this architecture is. It is how familiar it is. The Controller is an executive function. The Planner is a prefrontal working memory system. The Executor is the associative cortex doing pattern completion. The Critic is the anterior cingulate cortex, monitoring for error. Anthropic built, in software, a rough but recognizable sketch of the human brain's functional architecture — and discovered, as neuroscience discovered before them, that the "thinking" part is the smallest, most expensive, and most constrained piece.
"The 98% of non-AI code is not the frame around the painting. It is the museum — without which the painting cannot exist, cannot be seen, and cannot mean anything."
— Ron Jagannathan, DeepCog.ai2 Minsky's Prophecy
Marvin Minsky spent forty years at MIT arguing against the very thing the AI public most wants to believe: that intelligence is a substance — something you either have or you don't, something that can be poured into a machine once you find the right formula. His Society of Mind (1986) proposed the opposite. Intelligence is not a substance. It is a relationship — the emergent relationship among many small, unintelligent agents that each do exactly one thing and communicate through structured protocols.
Minsky called his agents "K-lines" — knowledge lines — and described how they activate in coalitions, with no single agent being "in charge" of cognition any more than a single neuron is in charge of thought. Criticized at the time for being too architectural and not mathematical enough, the book was, in retrospect, a design document for what Anthropic built four decades later.
The architecture Shrivastava described is, in the precise technical sense, a Society of Mind instantiated in code. The Controller, Planner, Executor, and Critic are Minsky's agents. The 98% of scaffolding — pipelines, validators, compliance layers — are the K-lines. And just as Minsky predicted, no single component "knows" what the system is doing. The cognition is in the coordination.
Key Insight: Minsky's most radical claim was that consciousness does not require a "self" — only a sufficiently complex society of interacting agents that model a self. If he is right, the question of whether Claude is "really" intelligent dissolves. It is like asking whether a parliament is "really" governing, when the parliament is all there is.
3 Penrose's Objection
Roger Penrose would not be satisfied by Minsky's answer — and this is where the intellectual tension becomes genuinely productive. In The Emperor's New Mind (1989), Penrose mounted a sustained argument against "strong computationalism" — the view that any sufficiently sophisticated computation constitutes, or could constitute, genuine understanding.
His argument drew on Gödel's incompleteness theorems. Gödel showed that within any consistent formal system powerful enough to express arithmetic, there exist true statements that cannot be proven within that system. Penrose's controversial move was to argue that human mathematicians can see the truth of Gödelian statements in a way no formal algorithm can. If human cognition transcends Turing-computable processes, then no amount of engineered scaffolding — however cleverly orchestrated — can produce the genuine article.
This forces us to be precise about what we mean when we call Claude Code "intelligent." If Penrose is right, Claude is an extraordinarily sophisticated simulation of reasoning — perhaps indistinguishable from reasoning in practice, but not equivalent in principle.
"The 2% that is 'AI' may be the most important 2% in the history of engineering. But it is still, in Penrose's terms, a very elaborate clock — not a mind."
— Roger Penrose's implicit challenge to the agentic paradigm4 What the 98% Actually Is
Setting aside the metaphysical question, what is the engineering function of the 98%? Why does so much non-neural code have to exist? The answer illuminates something important about intelligence itself. Minsky understood this: the agents in a Society of Mind do not produce cognition by thinking. They produce it by managing the conditions under which thinking can occur.
Consider the brain. The neocortex — the "thinking" part — constitutes roughly 76% of brain volume. But the brainstem, cerebellum, thalamus, hippocampus, and basal ganglia are not mere support tissue. They regulate attention, consolidate memory, gate sensory input, calibrate arousal, and manage timing of neural firing. Damage the thalamus, and the cortex goes dark. The scaffolding is not subordinate to intelligence. It is what makes intelligence possible.
| Dimension | Monolithic AI Model | Multi-Agent Architecture |
|---|---|---|
| Cognitive Model | Single large transformer — one system does everything | Distributed agents — specialized modules cooperate |
| Interpretability | Opaque — decisions emerge from billions of parameters | Auditable — each agent's contribution can be inspected |
| Error Handling | Single point of failure; errors propagate silently | Critic agent catches errors; distributed fault tolerance |
| Scalability | Bounded by model size and context window | Modular — agents can be added, swapped, or retrained |
| Minsky Analogy | A single neuron trying to be a brain | A society of agents — each dumb alone, intelligent together |
| Penrose Verdict | Sophisticated pattern matching; no claim to understanding | Sophisticated orchestrated pattern matching; same verdict |
DeepCog Principle: In healthcare AI, the 98% matters more than anywhere else. A Controller agent that misclassifies a symptom cluster sends the Executor in the wrong direction — regardless of how accurate the neural network inference is. At DeepCog.ai, our architecture explicitly separates clinical reasoning (the Executor) from clinical context management (the Controller) and safety validation (the Critic). The scaffolding is not an afterthought. It is the clinical governance layer.
5 Intelligence as Emergent Property
The Claude Code revelations converge with a growing consensus in cognitive science: intelligence is not a thing but a process. It is not located anywhere in the system — not in the neural weights, not in the orchestration logic — but emerges from their interaction.
Stephen Wolfram has argued something similar from a computational direction: that complexity itself cannot be shortcut. You cannot predict the output of a complex system faster than by running it. Intelligence may be the same — the only way to produce what looks like genuine reasoning is to actually run all the processes that constitute reasoning, including the 98% that doesn't look like AI at all.
Every line of orchestration code represents a human decision about how cognition should be structured. The Controller's task decomposition logic encodes a theory of problem-solving. The Critic's evaluation criteria encode a theory of correctness. The compliance layer encodes a theory of ethics. The engineers who wrote the scaffolding were doing philosophy — instantiating their beliefs about how a mind should work in executable form.
DeepCog Principle — Engineering as Ethics: If 98% of what makes an AI system behave the way it does is human-written code, then the primary locus of ethical responsibility is not the model — it is the engineers. The scaffold encodes values. To build AI ethically is first and foremost to engineer the scaffolding ethically: a Critic that catches what matters, a Controller that asks the right questions, a Planner that respects human autonomy.
6 From Chatbots to Cognitive Collaborators
Yudistira Ashadi's April 3 post framed the Claude Code architecture as evidence of a paradigm shift: from reactive AI — systems that answer when spoken to — to proactive agentic AI that plans, decomposes, executes, and self-evaluates within a set of constraints. A chatbot is a tool. An agentic system is a colleague. It maintains state across a task, anticipates the next step, catches its own mistakes, and can be held accountable for discrete decisions made by discrete agents. Multi-agent architecture makes AI more legible — which is a prerequisite for trust.
In healthcare, this distinction is existential. An agentic clinical reasoning system that misroutes a patient's care pathway — because the Controller encoded the wrong decomposition of "chest pain presenting at 3am" — causes harm through a chain of deterministic decisions, none of which looked wrong in isolation. The legibility of multi-agent architecture is not a luxury. It is a safety property.
✦ Conclusion: The Society of Code
Marvin Minsky would have read the Claude Code analysis with a particular satisfaction — not because he was right about the architecture (he was), but because the engineers arrived at it empirically, without starting from his theory. The Society of Mind is what you get when you try to build a system that actually works across the full complexity of open-ended human tasks. You discover, as Minsky did, that the problem requires a society.
Roger Penrose would read the same analysis with a different satisfaction. He would note that the Controller, Planner, Executor, and Critic are all Turing-computable processes. The system is more sophisticated than any previous AI — but it does not cross the threshold he cares about. Whether it ever can is not a question the leaked codebase answers. It may not be a question any codebase can answer.
- Intelligence, whether biological or artificial, is mostly infrastructure. The insight, the inference, the moment of apparent understanding — possible only because everything around them was engineered to make them possible.
- To build AI well is to take the 98% as seriously as the 2%. To regulate AI well is to understand that most behavior that matters happens in the scaffolding, not the weights.
- We need both Minsky and Penrose. The architect who showed minds can be engineered as societies, and the mathematician who reminded us that cleverness has limits that may be permanent.
- The age of agentic AI has arrived. It was, in the most important sense, predicted in 1986.