Connecting AI with the physical world: Three machine-economy directions Wintermute is bullish on

Author: Wintermute

Compiled by: Deep Tide TechFlow

**Deep Tide Brief: **Wintermute releases an industry manifesto: the battle over crypto’s infrastructure is over. The next battleground is not DeFi, but machine economics. When AI agents, warehouse robots, and automated experimental systems become economic actors, the foundational assumption of traditional finance—that “the other side is a human”—will fail completely. And the difference between the crypto rails—“the other side is code”—will turn from a flaw into a core advantage. Three directions worth watching: the agent economy layer, physical AI, and machine-driven discovery.

Old problems are dead—new problems are emerging

Crypto has already been around for more than ten years. L1 is already live, L2 follows right after; DeFi is mature, and stablecoins have become infrastructure. On every track—exchanges, lending, perpetual contracts, and prediction markets—every niche is crowded, and every obvious idea looks like it has already been done by someone.

So, what else can crypto build?

Many builders give up here. They’re wrong—not because the answer is “nothing,” but because the question itself is wrong.

For most of crypto’s history, the truly interesting questions were whether the rails could hold: whether settlement could happen in seconds, whether stablecoins could be transferred at scale, whether open networks could run under real load. These questions now have answers. Infrastructure can work—so the next interesting question is elsewhere.

What has truly changed is everything around the infrastructure. Models can act autonomously rather than just responding; robots learn from human videos rather than relying on hand-written code; open standards for agent payments and identity are starting to take shape. None of this is “crypto” by itself, but each item is pushing against the boundaries of the financial and trust infrastructure built for humans.

The question worth asking is no longer “what can crypto do,” but “what does this world need crypto to do.”

The answer is becoming clearer and clearer—machine economics.

Machines aren’t tools—they’re economic actors

When we say “machine economics,” we don’t mean machines as tools—used to send emails or write code. We mean machines as economic actors.

This shift may seem subtle, but the consequences are huge. Tools wait for instructions; actors hold context, make decisions, and execute transactions—autonomously acting in both the digital and physical worlds. Now models are good enough and cheap enough to do it at scale.

Real-world scenarios:

An agent books your flight, negotiates the price, pays the merchant, and handles refunds—end to end with no need for you to get involved.

A warehouse robot picks up tasks charged per item, charges itself, pays for compute itself, and routes its revenue to operators.

A research system autonomously designs experiments overnight, purchases reagents, and runs in a closed loop—no graduate students present.

Almost all of our existing financial and trust infrastructure assumes the other side is a person or a company—a party you can identify and hold accountable. That assumption collapses the moment the counterpart becomes an autonomous actor, and our existing rails for payment, identity, authorization, disputes, and settlement have never been built for this situation.

And this is exactly where crypto, fintech, AI, robotics, and quantum computing intersect.

Why now

Three recent shifts that, a few years ago, didn’t look very likely.

Models can act, not just answer

Models are no longer only answering questions. They can act autonomously, and the costs are low enough that they can run unattended. The unit cost of digital work is collapsing, making tasks that weren’t worth a human’s time before now feasible—and at the scale and dollar amounts that existing systems were never designed to handle.

Open standards are maturing

Stablecoins are now the real settlement rails. Protocols like x402, MPP, and AP2 provide agents with payment methods. Faster blockchain networks and faster fiat networks are converging in the middle. Open vision-language-action models allow robots to learn from human videos and simulations rather than relying on custom programming. Standards let builders compose instead of rebuild—and that’s precisely why progress is accelerating across every category.

Agents can run continuously

Unlike the tools we’re used to—tools that adapt to narrow, guided use cases—agents hold context and work autonomously over long periods without supervision. This changes the economics of automation, and changes the volume of activity any system must absorb.

Individually, none of these make an argument. Together, they do.

Crypto isn’t dead—it’s moved battlefields

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

The next wave of interesting companies won’t be crypto vs AI or crypto vs robots. The founders we’re most excited about aren’t choosing between these technologies—they’re stacking them.

You’re no longer building only within crypto. You’re building crypto + AI, crypto + robots, and crypto + autonomous science.

Traditional financial rails are built around human accountability: verifiable identities, intentions you can dispute, and people you can hold responsible when something goes wrong. Crypto rails are built around something different: code you can audit, on-chain records that anyone can read, and rules enforced by the network.

When the counterpart is an autonomous actor, that difference is no longer a flaw—it becomes critical. As machine-driven activity volume grows, rails built for crypto fit these needs better than rails designed for humans: open, programmable, permissionless, second-by-second settlement, and identity that doesn’t require intermediaries.

The opportunity for crypto builders isn’t competing with the crypto builders from the previous cycle; it’s becoming the foundational layer for the next wave of AI, robots, and physical autonomy.

And the biggest platforms are already sprinting. Over the past few months, Coinbase, Robinhood, and Binance have each rolled out agent trading infrastructure: operator wallets for agents and autonomous execution—Robinhood even built a brand-new chain specifically for it. This is no longer a niche crypto conversation; it’s happening on one of the world’s largest retail user platforms.

Failure modes today

The bet above is that permissionless, programmable rails are better suited to autonomous actors than rails designed for humans. That bet hasn’t yet been proven at scale, and two failure modes already show why more work is still needed:

Security: agent wallets have become an attack surface

In May 2026, an attacker used a Morse-code-style prompt injection to make Grok output a transfer instruction. The automated trading agent then executed it on-chain, moving roughly $150k to $200k, most of which was later recovered (SlowMist).

Accountability: who bears the consequences when AI systems fail

Even if AI, human reviewers, and governance votes all sign off, the question of who is responsible when a system involving AI fails still hasn’t been resolved. In February 2026, a prediction oracle bug in Moonwell’s AI-assisted smart contract code led to a $1.78 million bad-debt incident—no step in the review chain caught it (rekt.news).

Wintermute’s three directions

Right now, most activity is focused at the component layer: base models, robot hardware, stablecoins, and exchanges. These markets are crowded and well-funded—opportunities aren’t there.

The opportunity lies in what connects them: rails for trading, coordinating, and establishing trust between machines that don’t yet exist. Three directions stand out:

The agent economy layer

The difficult part isn’t whether agents can pay; it’s: when agents make mistakes, who holds permissions? Who bears fraud risk? And how can all of this reach merchants without requiring them to rebuild their checkout processes?

The form of the agent business is still being written: the authorization layer, agent identity, neutral routing across rails, and markets where agents purchase their own compute/data/access. The better teams here charge fees for authorization and risk reduction rather than taking a cut of payment value—making the business viable even before agent scale truly arrives.

Physical AI

Robots gain capabilities much faster than the speed at which economic scale is achieved. A model can now generalize across tasks and across different robot bodies; non-engineers only need to tell the robot what to do to redirect it. But robots still cannot pay for their own compute, charging, or maintenance, and they also cannot earn compensation for the work they do.

What’s missing isn’t hands—it’s wallets. We care more about structured scenarios—warehouses, logistics, and retail back offices—where economics are already viable and real deployments already exist, rather than home humanoid robots.

Machine-driven discovery

Software that orchestrates labs, designs automated experiments, and closes the loop between assumptions and results. Founders building autonomous layers for science are already selling into materials and drug-discovery labs. Quantum is the wildcard next to this direction: simulation and sensing could step-change what becomes discoverable, and post-quantum security is already a real requirement for the settlement layer. Hard to underwrite; the winner isn’t clear—yet there’s something here.

R[3]sidency × Construct accelerator

The infrastructure required for machine economics doesn’t exist yet. That’s where the work is—and where Wintermute is looking.

They want to support founders who can see new problems emerging between financial rails, autonomy, and trust—founders who can deliver products on existing rails while staying adaptable as standards evolve.

That’s the purpose of R[3]sidency × Construct: 8 teams, $300k per team, 12-week London residency, 30+ mentors, London and New York Demo Day. Run together with top partners: Fabric Ventures, Solana, and Coinbase.

If you’re building for a world where machines and people trade and operate in parallel—they want to support you.