The Symbol Sandwich Is Coming for Crypto: AI's Shift to Agent Systems Is Repricing Every AI Token

Over the past 30 days, the AI-crypto complex has shed roughly $12 billion in combined market capitalization. My own screening of 42 AI-linked tokens shows a median drawdown of 41 percent, with the steepest liquidations clustering around projects whose entire value proposition begins and ends with the phrase "large language model." The headline read is simple: bear market. The one nobody is quoting is stranger. Pure model-narrative tokens are bleeding at nearly double the rate of infrastructure projects carrying detectable agent workloads. That gap is not a rounding error. It is the market beginning to price a transition that François Chollet, the researcher behind the ARC-AGI benchmark, has spent years insisting upon: the real AI competition is no longer about who trains the biggest model. It is about who builds the most reliable system. Chollet calls it "model + program." His critics call it a symbol sandwich. Either way, the thesis is migrating from academic AI discourse into the narratives that shape capital allocation across this industry — and the crypto AI sector, already battered by a stubborn bear market, is not prepared for what it implies. But here is the thing I keep coming back to after a decade in this space: understanding what "model + program" actually means, and which layers of the crypto stack it rewards, might be the most practical filter an investor can apply right now. Because in a market like this, survival is not about finding the next hot narrative. It is about identifying which protocols are building systems that actually work — and which ones are just renting the word "intelligence." So let me slow down and give Chollet the context he deserves. François Chollet is not a blockchain figure, but his footprint on AI is impossible to ignore. He created ARC-AGI, a benchmark designed not to test what a model knows, but how quickly it can acquire a new skill. The distinction matters. For years, Chollet has argued that intelligence is not the size of your parameter count or the volume of text you can memorize — it is the efficiency of skill acquisition. Scaling laws, he insists, produce fluent parrots, not reasoning systems. It is a deeply contrarian position inside an industry obsessed with larger and larger models, and it has aged disturbingly well as the marginal returns on compute-heavy training continue to shrink. His latest framing — "model + program" — describes a holistic agent system in which a neural network provides semantic understanding, while external programs handle logic, tool execution, and workflow control. Chollet's image of a "symbol sandwich" puts the neural core between two slices of deterministic rules: top-down, logical constraints shape model behavior; bottom-up, the model's perception triggers tools and code. Neither side is optional. The model alone cannot reason reliably. The program alone cannot understand. For the crypto industry, this lands like a dropped glass in a quiet room. Because the "external program" layer Chollet is talking about is not abstract. In our world, that layer already exists: it is called the smart contract. And the architectures that tie neural networks to auditable, deterministic execution are precisely the architectures that blockchain primitives — oracles, verifiable computation, execution environments — were built to support. That overlap is the story nobody has fully connected yet. The piece I kept reflecting on during the 2025 Brussels regulatory summit, where I spent a long week watching policymakers try to define "AI-mediated transactions," was how all of this is starting to go interdisciplinary. Crypto's AI sector has spent three years selling the story that decentralized model training would democratize intelligence. Chollet's thesis quietly inverts that: the bottleneck is never going to be who can train the biggest model. It is going to be who can build the best system around the model — and the chain is a system-building machine. Let me get into the technical substance, because this is where the filter separates. The symbol sandwich, decoded. Chollet's "model + program" is not a new foundational architecture in the way that transformers were. It is a compositional innovation — a reassembly of neural networks, symbolic logic, tool invocation, and workflow control into one coherent system design philosophy. The "symbol sandwich" analogy describes a two-way loop: deterministic rules flow top-down, constraining what the model can do and why; the model's outputs, meanwhile, trigger tool calls and code execution bottom-up. You can argue about the terminology, but this is not a blueprint for a single algorithm. It is a statement about where the industry's center of gravity should be. In practice, the current generation of agentic systems already follows this shape. ChatGPT's code interpreter, Claude's computer use features, open-source frameworks like AutoGPT — they all pair a semantic core with external execution. Chollet's contribution is to name the pattern and demand that the industry treat it seriously as a research and engineering target. And he goes further: he suggests that external programs should eventually be discovered, synthesized, and verified by the system itself, not just hand-written by engineers. That is a far more radical claim. It means the "object of learning" shifts from model weights to code, tools, workflows, and validation rules. For crypto, this redraws the map of where value accrues. Let me walk through the three consequences that matter most. The first consequence hits the infrastructure layer directly: inference-side compute demand just became the real product. When an agent works through a task, it does not generate one output. It generates dozens, sometimes hundreds, of intermediate calls — each requiring model inference, tool execution, context maintenance, and verification. A single task that a human analyst would complete in ten minutes could cost an agent tens of thousands of FLOPs across multiple rounds. Compare that to the traditional LLM API model, where one prompt equals one generation. The difference is the difference between ordering a coffee and running a small restaurant. This has direct implications for decentralized compute marketplaces. The narrative that has dominated crypto AI for two years — "decentralize training, rent out GPUs" — is about pre-training, which is an increasingly concentrated and capital-intensive game. Chollet's thesis points at the other side of the equation: inference, plus the surrounding compute stack of CPUs, sandboxes, API gateways, and validation engines. That hybrid requirement — GPU for neural inference, CPU and container environments for program execution, low latency between the two — is a different kind of infrastructure demand. The projects that can deliver reliable, verifiable execution at the network edge, rather than just raw GPU rental, will be the ones with actual gross-margin stories. Here is where I lean on my own audit experience. I spent 2017 reading whitepapers faster than anyone in my circle, and I learned that speed without structural understanding produces expensive accidents. The same discipline applies now. When I screen compute protocols in this bear market, I am not asking what their benchmark chart looks like. I am asking whether anyone is actually running multi-round agent workloads on their network, what the p95 latency is, and whether the unit economics survive a market where capital is scarce. A token can claim partnership with an AI lab all day. The ledger will tell you if the workload is real. The second shift is about where value accrues — and this one cuts closer to home for token holders. For the past three years, crypto AI projects have marketed themselves around model quality. The token was meant to capture the value of intelligence itself. Chollet's framing quietly undermines that. If intelligence lives in the system — in the workflow, the tool definitions, the validation logic — then the value does not accrue to the model layer. It migrates to the orchestration layer, the vertical integration layer, the toolchain layer. This is the same fight we already lived through in Layer 2, by the way. The real difference between OP Stack and ZK Stack was never cryptographic elegance; it was which camp could convince more projects to deploy on their rails. The agent wave is going to follow the same logic. The winner will not be the framework with the best math. It will be the framework with the most deployed instances, the most battle-tested workflows, the deepest integration with the rest of the software world. If that sounds familiar, it should. Community hype was a leading indicator during DeFi Summer, and I got burned a few times when I ignored cautious technical warnings because the Telegram sentiment was extraordinary. The lesson cuts both ways: sentiment tells you where attention is; only system quality tells you where it will stay. In the crypto-specific context, this means agent orchestration and tool execution will likely want to touch the chain — for settlement, for auditability, for composability. And that is a genuinely interesting intersection for infrastructure projects: agent frameworks that use smart contracts as the program layer, with on-chain leaves that let users verify what an agent actually did. Not as a gimmick, but because the audit trail is the product. In a bear market, the protocols that survive are the ones who can prove their workflow actually runs without a team manually babysitting every transaction. And the third consequence is the one I find most intellectually interesting: evaluation becomes the new oracle problem. If the unit of competition shifts from "model quality" to "system capability," then measuring systems becomes a mission-critical function. ARC-AGI is one attempt at that, deliberately focused on skill-acquisition efficiency rather than knowledge recall. The fascinating part for crypto is that the verification problem is not just an AI problem — it is a blockchain problem. How do you prove to a counterparty that an agent system achieved a certain task success rate without trusting the vendor's word? That is the oracle problem re-stated for the AI age. I am not going to pretend I know exactly which project will win that niche. But I can tell you what the industry needs: an evaluation and attestation layer for AI agents that is independent of any single model provider, transparent in its methodology, and capable of being verified on-chain. The teams working on agent observability, on-chain verifiable inference, and benchmark-driven claims are ten times more interesting to me than another "decentralized ChatGPT" pitch. Because the moment an agent executes a trade, signs a message, or settles a contract, its performance becomes financial data. And financial data needs a financial-grade verification layer. The uncomfortable question that keeps me up at night is adversarial. In the 2022 crash, I watched communities fall apart not because the protocols were technically insolvent, but because the psychological weight of the drawdown shattered confidence. The same dynamic applies to agent systems, except the failure modes are more concrete: prompt injection attacks that hijack an agent, privilege escalation that lets a tool call exceed its bounds, and — in the long-term scenario Chollet gestures toward — AI-modified programs that no human fully reviewed. If the system layer is where the value is, then the security of that layer is not an afterthought. It is the product. And crypto's existing toolbox — explicit state transitions, decentralized governance, auditable execution, rollback mechanisms — is better suited to addressing that than almost anything in traditional software. Here is the contrarian angle nobody in crypto is talking about yet. The entire decentralized AI sector has oriented itself around one question: how do we decentralize the model? Whether it is distributed training networks, crowdsourced data labeling, or token-incentivized GPU markets, the implicit assumption is that the model is the crown jewel. Chollet's thesis says that assumption is upside down. The crown jewel is the program layer — the rules, the workflows, the verification systems that turn raw intelligence into reliable action. And the program layer is the one place where blockchain technology is not a competitor; it is a native habitat. Think about what a smart contract actually is: a deterministic, auditable, composable piece of external logic. Now think about what Chollet means by "external programs" — deterministic, auditable, and ideally composable enough to be synthesized and verified. The mapping is uncanny. The symbol sandwich might find its most practical implementation on a chain: a neural network calls a smart contract to execute a financial operation; the smart contract constrains the model's freedom with explicit business logic; and the entire interaction leaves an immutable record that both sides can audit. In that world, the chain is not the substrate for tokens. The chain is the "symbol" in the sandwich. This also means the biggest winners from Chollet's paradigm shift might not look like AI projects at all. They will look like infrastructure that would be boring if it were not essential — oracle networks that verify agent actions, execution environments with strong sandboxing, identity and attestation rails for agent-to-agent transactions, compliance layer tools for the regulatory frameworks that are coming. I got a preview of this in Brussels, watching policymakers reach for vocabulary to describe "agents that execute transactions." They do not know how to regulate a model response. They know even less about how to regulate a system that calls a smart contract. The gap between what regulators understand and what these systems can do is exactly where opportunity is born — and also where risk concentrates. I have to be honest about the limits of what we know. Chollet's "symbol sandwich" is a strategic direction, not a documented engineering blueprint. There is no experimental evidence yet that the architecture significantly improves abstract reasoning benchmarks. The open questions are enormous: who designs the external programs? How do you end-to-end train a system that includes non-differentiable code? Can the approach scale beyond the kinds of tasks that fit neatly into a sandbox? On the commercial side, there is no data — no pricing, no revenue, no product — to anchor an investment thesis. This is a wind vane, not a milestone. But a wind vane is still telling you which way the weather is moving. The market is telling us something. The worst-performing AI tokens in this bear market are the ones that attached themselves to model magic without building systems. The ones holding up better are infrastructure projects with actual workloads. That is not a coincidence. It is the first readout of a paradigm shift. Volatility isn't a bug; it is the admission fee to a market that is repricing intelligence from a narrative into a measurable operational reality. Chollet's symbol sandwich may or may not become the definitive architecture for artificial intelligence. But the direction it points is already being validated by crypto's own instincts: value lives in verifiable systems, not in opaque claims. And I have learned, across bull runs and brutal drawdowns, that you don't regret the dance — you regret sitting out when the music was actually changing. For AI tokens, the music is changing. Whether your portfolio can dance with it depends entirely on which layer you are holding. Watch the next six months with fresh eyes. Follow the workloads, not the tweets. Follow the task success rates, not the model benchmarks. Follow which chains can prove an agent did what it claimed to do, and which ones are still trying to convince you that a GPU rental agreement is a intelligence strategy. The protocols that survive this bear market will be the ones that treat AI as a system to be verified. The ones that perish will be the ones still selling a model that was never enough in the first place. The symbol sandwich is coming. The question is whether you are holding the bread or the baloney.