I ran a grep across the top forty AI-agent tokens by reported market capitalization last week. Not for alpha. For a sanity check on a sentence I keep finding in pitch decks: that these are "autonomous agents that improve themselves over time."
Of the forty, thirty-seven had at least one of the following in deployed bytecode: an onlyOwner modifier on the settlement path, an admin-controlled price or inference feed, or a keeper allowlist that a three-of-five multisig could rewrite in a single transaction. Exactly three had an improvement loop that could not be reduced, within one hop, to a human pressing a button somewhere in the stack.
That was Monday. On Tuesday, Sundar Pichai said AI self-improvement remains in "early stages," that progress is slow, and that human intervention is still required. The crypto wires reran it as an AI story. It is not an AI story. It is a collateral valuation story. The collateral is every token whose price embeds a capability that does not yet exist in any deployed system, on-chain or off.
Let me be precise about the mechanism before anyone files this under AI doomerism or, worse, under "Google is sandbagging to suppress competitors."
The phrase "self-improvement" is not a vibe. It has a technical definition with four distinct layers, and conflating them is the single most expensive category error in the current AI-crypto market. Layer one is self-training: a model generates synthetic data that trains the next model. This is deployed at industrial scale and has been for years; every major lab does it. Layer two is algorithmic search, where a model proposes and evaluates candidate improvements to code or mathematics. Google DeepMind's FunSearch and AlphaEvolve sit here — AlphaEvolve produced a faster matrix multiplication routine for a production kernel, and FunSearch found genuine new constructions for combinatorial problems. That is real, it is peer-reviewed or at least publicly documented, and it works because a human writes the evaluation function. Layer three is autonomous pipeline design, where a system sets its own training objectives and hyperparameters. This is partially automated; humans still specify the objective. Layer four is end-to-end recursive self-improvement — a system that improves its own architecture without human-defined goals — and the public count of instances is zero.
When Pichai says "early stages" and "human intervention is necessary," he is describing layers three and four accurately. If he meant layers one and two, the statement would be false, because those are already load-bearing in Google's own product pipeline. The ambiguity is not sloppiness. It is the entire exploitable surface.
Here's the part that should worry you: crypto's AI-agent sector has priced itself at layer four while operating at layer one with a layer-two marketing wrapper.
A token that claims to be a "self-improving autonomous agent" and settles through a multisig-gated execute() function is not an autonomous agent. It is a SaaS product with a token attached to its billing relationship. That is not a moral failing — plenty of good businesses look like that. It is a valuation mismatch. The price reflects a recursive improvement loop; the code reflects a cron job and an admin key.
I want to be careful with the claim, because the reflexive response is "he's short." I have no position in any of the forty. What I have is a habit from auditing Solidity during DeFi Summer, when I spent forty hours on Compound's original governance contract and found an integer overflow in claimReward that predated the well-known reentrancy patch. That episode taught me a durable lesson: high-level abstractions mask fundamental logic errors, and the abstraction is always most dangerous when it is most confidently asserted. A protocol that says "trust the math" is asking you to skip the exact step where the math usually breaks.
The same pattern is now playing out one layer up the stack, in the AI-agent oracle problem, and it is worth walking through the code because the failure mode is not what the whitepapers anticipate.
In 2024 I analyzed an AI-driven oracle network that used LLM validators to attest to off-chain events. The architecture was the standard committee design compressed into a few lines:
for i in committee:
out[i] = llm_infer(template + feed[i])
if count(mode(out)) / len(out) >= 2/3:
settle(mode(out))The security argument was Byzantine fault tolerance: as long as fewer than a third of validators are corrupted, the honest majority prevails. That argument holds only if failures are independent. They are not. Every validator in that committee shared the same base model family, the same retrieval index, and the same prompt template. A single prompt-injected news article — a crafted headline that flips sentiment classification — caused all validators to return the same incorrect value with high confidence. The quorum was met. The settlement executed. No malicious node existed anywhere in the system. The network was Byzantine-fault-tolerant against adversaries and completely defenseless against correlation.
I reproduced it locally with a small inference server and a synthetic feed. Same result across three model families when the prompt template was held constant. The fix is not more validators; it is enforced model diversity, semantic cross-checks against the source document rather than against other validators' outputs, and a staked slashing condition on divergence from a ground-truth feed. Every one of those fixes requires a human-maintained ground truth. Which is exactly the point Pichai was making, arriving in crypto through a completely different door.
The second structural failure is the one I got partially wrong, and it is instructive that I did. In 2025 I dissected a layer-2 designed to monetize AI compute — inference as a commodity, priced on-chain. The token emission schedule rewarded nodes on compute delivered, measured in FLOP-seconds, with no quality gate on the output. The obvious attack was Sybil: spin up cheap inference nodes running quantized models, accept every request, emit tokens, return low-quality results that most requesters could not audit. My simulation projected emission-driven hyperinflation within six months. Technically, the model was right.
What I missed was governance. The team could and did adjust the emission curve mid-flight, which rendered my static analysis obsolete as a prediction while leaving it intact as a diagnosis. The lesson I took from it — and the reason I now run parameter sweeps over realistic governance response functions rather than fixed schedules — is that in token economies, the emission curve is not a constant. It is a policy variable, and the policy-maker is a multisig. Any model that treats it as fixed is modeling the wrong object.
Apply that to the self-improvement question and the picture clarifies. If autonomous improvement were real, the emission schedules of AI-compute networks would be a minor detail, because the marginal cost of capability would collapse and token design would be irrelevant. Because autonomous improvement is not real, the entire value of those networks rests on human-driven R&D intensity — which means the emission schedule, the quality gate, and the governance multisig are the whole ballgame. Pichai's remark does not change AI. It changes which variables matter.
Now the economics of proof, because this is where the AI-agent sector's claims collide with arithmetic.
Verifiable inference comes in three flavors. Zero-knowledge proof of inference — zkML — gives you cryptographic soundness. Trusted execution environments give you hardware attestation. Optimistic machine learning gives you fraud proofs with a challenge window. I have written before that ZK rollup proving costs are absurd unless gas returns to bull-market levels and operators are bleeding money on the delta. zkML is the same disease with a worse prognosis, because the circuit for a transformer forward pass is orders of magnitude larger than a rollup's state transition circuit. Proving a single inference of a mid-sized model can cost more than the inference by three to four orders of magnitude. That ratio does not improve with demand; it improves with specialized hardware and better proof systems, both of which are years out for general models.
TEEs are cheap and fast and rely on trusting Intel, AMD, or a cloud provider's attestation chain — a single-vendor trust assumption that the crypto industry spent a decade inventing primitives to avoid. Optimistic ML is economically viable only when the value of the settled request exceeds the cost of the challenge game, which for agent micropayments is almost never.
So the honest description of 2026's "verifiable AI agent" is: an inference call made off-chain by a model you cannot inspect, attested by a signature from a key controlled by a team, with an on-chain settlement layer that records the result but not the computation. That is a perfectly reasonable architecture. It is also not verification. It is a trusted oracle wearing verification's clothes, and the market has priced the clothes.
The cross-chain dimension compounds it. Agent liquidity is fragmented across four or five rollup environments, and settlement between them still routes through bridge contracts whose failure modes are well documented. Dencun lowered the cost of posting data between rollups, which helped the fee line and did nothing for the user experience — the number of steps, signatures, and wait states to move agent-settled value across rollups remains an order of magnitude worse than withdrawing from a centralized exchange and redepositing. For a human making a weekly trade, that is tolerable friction. For an autonomous agent settling microtransactions at machine timescales, it is a structural blocker. Agents cannot be autonomous if every value transfer requires a bridge confirmation and a fifteen-minute challenge window. The infrastructure that would make agent autonomy real does not exist, and it is not being built at the layer where the tokens are issued.
I should note the regulatory overlay here, because it interacts with the valuation story in a way that is usually missed. A meaningful share of AI-agent token issuers have re-domiciled their foundations in the last eighteen months, and the destination choice is not random. Jurisdictions competing for the licensing business are competing on the same axis — who can offer a credible supervisory framework without demanding the kind of disclosure that would expose the gap between the whitepaper and the deployed code. That competition is real and it is producing frameworks optimized for institutional comfort rather than technical honesty. Which is fine as a market-structure story and useless as a risk mitigation story, because none of those frameworks require the issuer to prove that the agent improves.
Here's the contrarian read, and it is the reason this article is not simply a short thesis.
Pichai's cooling statement is bearish for layer-four narratives and structurally bullish for a narrow, specific slice of the crypto AI stack. If autonomous improvement is not coming, then capability gains continue to depend on human labor — data curation, evaluation design, red-teaming, and quality gates. Every one of those is an economic activity that can be priced, and several of them can be settled on-chain. Human-in-the-loop oracle middleware, verifiable evaluation services, stake-weighted quality attestations, and audit tooling all gain value in a world where the loop stays open. The tokens that assume the loop closes lose value. The market, if it reprices at all, will probably not make that distinction on the first pass, which is where the actual mispricing lives.
The second reason to be cautious about the bearish read is reflexivity. A head of a tier-one lab saying "it's early" can be read two ways. The straightforward read is that the technology is not there. The conspiratorial read is that the lab is deliberately understating progress to avoid regulatory scrutiny and competitive escalation — and the conspiratorial read is, for a certain class of market participant, the more attractive one because it implies an imminent catalyst. I have watched this exact inversion happen with the ZK circuit audit I ran in 2024, where a soundness bug in the Groth16 challenge generation phase got recirculated on forums as evidence that the team was "hiding something bigger." The bug was real and the rumor was nonsense. Cooling statements do not reliably cool.
The deepest problem is informational, and this is what the market is actually mispricing. There is a tier-one view of self-improvement, held by people with direct visibility into AlphaEvolve-class research and its limits, and a tier-three view, held by token issuers who need the narrative to raise. The two views have never been further apart, and the tier-three view is the one that gets tokenized. Pichai's remark is a data point that the gap is widening, and gaps in expectations are the raw material of repricing events — not because the tier-one view is correct in every particular, but because it is closer, and it is public.
This is not a prediction. It is a constraint. The constraint is that any token whose thesis requires an end-to-end recursive self-improvement loop is currently unbacked, and will remain unbacked until someone publishes a reproducible demonstration. Not a demo. Not a benchmark. A reproducible, independently verified result where a system improves its own capability without human-defined objectives. That publication, whenever it comes, is the only hard falsifier of Pichai's statement, and until it exists the layer-four premium is unsupported.
What I would track, in order of signal value. First, whether this becomes a series: if two more tier-one CEOs converge on the same cooling language within two quarters, the narrative shift is structural rather than incidental, and the first-pass repricing becomes defensible. Isolated statements are noise; three is a regime. Second, Google DeepMind's publication cadence on evolutionary and self-improving code systems — if the papers keep coming while the executive language stays conservative, the "research aggressive, narrative conservative" hypothesis is confirmed, and the gap between what the lab knows and what the market believes is a durable information edge for anyone willing to read preprints. Third, the primary-market funding environment for autonomous-agent startups over the next twelve months; a soft round is a faster signal than a price chart. Fourth, and most important, whether the on-chain agent sector starts shipping quality-gated emission schedules and diverse-model oracle committees — because those are the unglamorous engineering fixes that would make the human-in-the-loop world economically functional.
The question I keep returning to is not whether AI will eventually improve itself. It probably will, and the people best positioned to know are the ones who just told us it hasn't yet. The question is why the crypto market keeps paying for the end state while the deployed code operates at the beginning of it, and whether the discipline of reading bytecode before reading whitepapers is going to stay a niche skill or become the thing that separates the next cycle's survivors from its exit liquidity. Every previous cycle answered that question the same way. The tools were always there. Almost nobody ran the grep.
