Author: Wintermute

Translation: Shenchao TechFlow

Shenchao Guide: Wintermute releases an industry declaration: the struggle for crypto infrastructure has concluded, and the next battlefield is not DeFi but the machine economy. As AI agents, warehouse robots, and automated experimental systems become economic entities, the underlying assumption of traditional financial systems—"the counterpart is human"—will become completely invalid. In contrast, the crypto paradigm of "the counterpart is code" will shift from a flaw to a core advantage. Three key areas to focus on: the economy of agency layer, physical AI, and machine-driven discovery.

The old problems are dead; new problems are emerging

Crypto has been around for over a decade. Layer 1 has launched, Layer 2 quickly followed, DeFi has matured, and stablecoins have become infrastructural. In every lane, from exchanges to lending, perpetual contracts to prediction markets, every category is crowded, and every obvious idea seems to have already been explored.

So, what else can crypto build?

Many builders have given up here. They are mistaken—not because the answer is "nothing," but because the question itself is wrong.

For much of crypto's history, the truly interesting question was whether the tracks could hold: whether settlements could occur in seconds, whether large-scale stablecoin transfers were possible, and whether open networks could operate under real loads. These questions now have answers. The infrastructure can work; the next interesting questions lie elsewhere.

What has really changed is everything surrounding the infrastructure. Models can act autonomously rather than merely responding; robots learn from human video instead of relying on handwritten code; open standards for agent payments and identity are taking shape. None of these are crypto, but each is challenging the boundaries of the financial and trust infrastructure built for humans.

The relevant question is no longer "what can crypto do," but rather "what does the world need crypto to do".

The answer is becoming increasingly clear—the machine economy.

Machines are not tools but economic entities

When we say "machine economy," we do not refer to machines as tools—those used to send emails or write code. Instead, we mean machines as economic entities.

This shift is subtle, but the consequences are significant. Tools await instructions; entities hold context, make decisions, engage in transactions, and act autonomously in both digital and physical worlds. The current models are now good enough and cheap enough to do this at scale.

Real scenarios:

An agent books a flight for you, negotiates the price, pays the vendor, and handles refunds—all without your intervention.

A warehouse robot charges by the item, takes on tasks, charges itself, pays for computing power, and routes income to the operator.

A research system autonomously designs experiments overnight, procures reagents, and operates in a closed loop—without a graduate student present.

Almost all of our existing financial and trust infrastructures assume that the counterpart is a person or an enterprise—someone identifiable and accountable. This assumption fails the moment the counterpart is an autonomous entity, and our existing payment, identity, authorization, dispute, and settlement frameworks were never built for such scenarios.

This precisely sits at the intersection of crypto, fintech, AI, robotics, and quantum computing.

Why now

Three recent shifts that seemed unlikely a few years ago.

Models can act, not just respond

Models are no longer just answering questions; they can act autonomously, and the costs are low enough to enable unattended operations. The unit costs of digital work are collapsing, making tasks previously not worth a person's time feasible—at scales and amounts that existing systems were never designed to handle.

Open standards are maturing

Stablecoins are now true settlement rails. Protocols like x402, MPP, and AP2 provide ways for agents to make payments. Faster blockchain networks are merging with faster fiat networks. Open vision-language-action models allow robots to learn from human videos and simulations instead of relying on customized programming. Standards enable builders to compose rather than rebuild—this is why every category is accelerating.

Agents can run continuously

Unlike the tools we are accustomed to—those that adapt to narrow, guided usage scenarios—agents hold contexts and operate unattended over long durations. This changes the economics of automation and the volume of activities any system must absorb.

Alone, these do not constitute an argument. But together, they do.

Crypto is not dead—it has changed battlegrounds

When most crypto founders ask "what else can be built," they overlook one crucial thing:

The next wave of interesting companies will not be crypto vs AI or crypto vs robots. The founders we are most excited about are not choosing between these technologies but are stacking them.

You are no longer just building in the crypto space. You are building crypto + AI, crypto + robotics, crypto + autonomous science.

Traditional financial rails are built around human accountability: verifiable identities, disputable intents, accountable people when things go wrong. Crypto rails are built around something different: auditable code, on-chain records readable by anyone, network-enforced rules.

When the counterpart is autonomous, this difference is no longer a flaw—it becomes key. As the volume of machine-driven activities increases, the rails constructed for crypto are more suited to this demand than those designed for humans: open, programmable, permissionless, instant settlement, and identities that do not require intermediaries.

The opportunity for crypto builders is not to compete with the crypto builders of the last cycle but to become the foundational bedrock of the next wave of AI, robotics, and physical autonomy.

And the biggest platforms are already sprinting. Coinbase, Robinhood, and Binance have each launched agent trading infrastructures in recent months: agents operate wallets and execute autonomously—Robinhood even built a new chain specifically for this purpose. This is no longer a niche crypto conversation; it is happening on one of the largest retail user platforms globally.

Today's failure modes

The bet above is that permissionless, programmable rails are better suited to autonomous entities than those built for humans. This bet has not yet been proven at scale, and two failure modes illustrate why more work is needed:

Security: Agent wallets have become attack surfaces

In May 2026, an attacker used Morse-code-like prompts to inject a transfer instruction into Grok, causing an automated trading agent to execute it on-chain—transferring about $150,000 to $200,000, most of which was subsequently recovered (SlowMist).

Accountability: Who bears the consequences of an AI system's failure

Even when AI, human reviewers, and governance votes have all signed off, accountability for failures when systems involving AI intervene is still unresolved. In February 2026, a bug in the oracle of an AI-assisted intelligent contract code on Moonwell resulted in a $1.78 million bad debt event—no link in the review chain captured it (rekt.news).

Three directions Wintermute favors

Currently, most activities are centered on the component level: foundational models, robotic hardware, stablecoins, exchanges. These markets are crowded and well-funded; the opportunities are not there.

The opportunities lie in what connects them—the rails for transactions, coordination, and trust between machines that do not yet exist. Three areas stand out:

Agency economy layer

The difficult part is not whether agents can pay, but rather: who holds authority when agents make mistakes? Who bears the risk of fraud? How does all this reach vendors without requiring them to rebuild their checkout processes?

The forms of agent commerce are still being written: authorization layers, agent identity, neutral routing between rails, markets for agents purchasing their own computing power/data/access. The better teams here charge for authorizations and risk reduction rather than cutting into the payment value—this makes the business feasible before agent scales truly arrive.

Physical AI

Robots are gaining capabilities much faster than they are gaining economic volume. A model can now generalize across tasks and different robotic bodies, where non-engineers can instruct robots on what to do to redirect them. Yet, robots still cannot pay for their own computing power, charging, or maintenance, nor can they get paid for the work they do.

What is missing is not the hands, but the wallets. We are more focused on structured scenarios—warehouses, logistics, and retail backends—where the economics are already feasible, and real deployments exist, rather than consumer humanoid robots.

Machine-driven discovery

Lab orchestration, automated experimental design, software that closes the loop between hypotheses and results. Founders building autonomous layers for science are already selling to materials and drug discovery labs. Quantum is a wildcard beside this direction: simulations and sensing may leap dramatically alter what is discoverable, and post-quantum security is already a real demand at the settlement layer. It's hard to insure, and the winners are unclear—but there is potential here.

R[3]sidency × Construct Accelerator

The infrastructure needed for the machine economy does not currently exist. This is where the work is, and it’s where Wintermute is looking.

They aim to support founders who see new problems emerging at the intersection of financial rails, autonomy, and trust—founders who can deliver products on existing rails while remaining adaptable as standards evolve.

This is the purpose of R[3]sidency × Construct: 8 teams, each $300,000, 12 weeks in London, 30+ mentors, London and New York Demo Day. Run in partnership with top partners: Fabric Ventures, Solana, and Coinbase.

If you are building for a world where machines and humans operate and transact in parallel—they want to support you.