Wintermute says crypto’s long-running fight over whether the rails can work is no longer the most interesting question. In a new industry thesis, the firm argues that base-layer and settlement infrastructure has matured enough that the next buildout will be driven by the needs of the machine economy rather than another round of purely internal crypto competition.
The report says Layer 1 networks are live, Layer 2 systems followed, DeFi has matured, and stablecoins have become infrastructure. In markets such as exchanges, lending, perpetuals and prediction markets, categories already look crowded. That has led many builders to ask what is still left to build in crypto. Wintermute’s answer is that the framing itself is wrong. The more useful question now is what this changing world will need crypto to do.
Its answer is increasingly clear, according to the report: machine economies, where the counterparties are not always people or companies, but autonomous systems.
Machines as actors, not tools
Wintermute draws a distinction between machines used as tools and machines acting as economic participants. It is not talking about software that helps send emails or write code. It is talking about systems that hold context, make decisions, transact, and act autonomously across digital and physical environments.
The report gives a few examples. An agent could book a flight, negotiate a price, pay the merchant and process the refund without user intervention. A warehouse robot could take work on a per-item basis, recharge itself, pay for compute, and route earnings back to an operator. An automated research system could design experiments overnight, buy reagents and run a closed loop without a graduate student standing by.
That shift may look subtle, but Wintermute says the consequences are large. Existing financial and trust systems are built around the assumption that the party on the other side is a human or a business entity that can be identified and held accountable. Once the counterparty becomes an autonomous actor, that assumption starts to break. Payments, identity, authorization, dispute processes and settlement rails were not built for that setup.
Why now
Wintermute points to three recent changes that, in its view, make the timing different from a few years ago.
Models can act, not just answer
Models are no longer limited to responding to prompts, the report says. They can now take actions on their own, and the cost has dropped enough to let them operate unattended. As the unit cost of digital labor falls, tasks that were not worth human time begin to make economic sense. They can also happen at volumes and transaction sizes that existing systems were never designed to absorb.
Open standards are taking shape
Wintermute says stablecoins now function as real settlement rails, while protocols such as x402, MPP and AP2 give agents a way to pay. Faster blockchain networks and faster fiat networks are moving closer together. At the same time, open vision-language-action models are letting robots learn from human video and simulation instead of depending entirely on custom programming. Standards matter because builders can compose instead of rebuild, which is one reason categories move faster once the interfaces start to settle.
Agents can run over longer periods
Unlike traditional tools that fit narrow, guided use cases, agents can retain context and work for long stretches without supervision. Wintermute says that changes the economics of automation and raises the amount of activity any supporting system must handle. None of those shifts alone is enough to make the full case, the report says. Together, they are.
Crypto’s role changes with the counterparty
Wintermute argues that the next interesting companies will not be built by choosing crypto over AI or crypto over robotics. The stronger founders, in its view, are stacking those technologies together. The buildout is no longer just in crypto. It sits in crypto plus AI, crypto plus robotics, and crypto plus autonomous science.
That matters because traditional financial rails were built around human accountability: identities that can be verified, intent that can be disputed, and a person or entity that can be pursued when something goes wrong. Crypto rails were built around something else: auditable code, on-chain records that anyone can read, and rules enforced by the network.
When the party on the other side is autonomous, Wintermute says, that difference stops looking like a weakness and starts looking useful. As machine-driven activity grows, open, programmable, permissionless rails with second-level settlement and identity systems that do not require intermediaries may be a better fit than systems designed for people.
The opportunity for crypto builders, according to the report, is not simply to compete with the last cycle’s crypto startups. It is to become the infrastructure layer for the next generation of AI, robotics and physical autonomy.
Wintermute also notes that larger platforms are already moving. Coinbase, Robinhood and Binance have each launched agent trading infrastructure over the past few months to support agent-operated wallets and autonomous execution. Robinhood, the report says, has even built a new chain for that effort. Wintermute frames that as evidence that the idea is no longer a niche crypto conversation and has reached one of the world’s biggest retail user platforms.
Two failure modes already visible
Wintermute says its broader bet is that permissionless, programmable rails will fit autonomous actors better than systems designed around people. It also says that bet has not yet been proven at scale. Two existing failure modes show why more work is still needed.
Security: agent wallets are becoming an attack surface
The report cites a May 2026 case in which an attacker used Morse-code-style prompt injection to get Grok to output a transfer instruction. An automated trading agent then executed it on-chain, moving about $150,000 to $200,000, with most of the funds later recovered. Wintermute attributes that case to SlowMist.
Liability: it is still unclear who absorbs the damage
Even when AI systems, human reviewers and governance votes have all signed off, Wintermute says responsibility remains unresolved when an AI-involved system fails. It points to a February 2026 incident on Moonwell, where an oracle bug in AI-assisted smart contract code caused $1.78 million in bad debt. According to the report, no step in the review chain caught the issue. That example is attributed to rekt.news.
Three areas Wintermute is watching
Most market activity today is clustered at the component level, the report says, including foundation models, robot hardware, stablecoins and exchanges. Those markets are crowded and already well funded. Wintermute says the more interesting opportunity lies in the connective layer: rails for trade, coordination and trust between machines that do not yet fully exist at scale.
1. The agent economy layer
Wintermute says the hard part is not whether an agent can pay. The hard part is who holds authority when an agent makes a mistake, who takes fraud risk, and how these systems can reach merchants without asking them to rebuild checkout flows. The commercial structure around agents is still being written, in its view. That could include authorization layers, agent identity, neutral routing across rails, and markets where agents buy their own compute, data and access.
The firm says the stronger teams here are more likely to charge for authorization and risk reduction than to take a cut of payment volume. That matters because such businesses could work before agent scale fully arrives.
2. Physical AI
Wintermute argues that robots are gaining capability faster than they are gaining economic scale. A single model can now generalize across tasks and across different robot bodies, and non-engineers can redirect a robot by telling it what to do. But robots still cannot pay for their own compute, charging or maintenance, and they cannot directly collect payment for the work they perform.
In Wintermute’s phrasing, what is missing is not hands but wallets. The report is more interested in structured environments such as warehouses, logistics and retail backrooms than in household humanoids, because the economics are already clearer and real deployments already exist in those settings.
3. Machine-driven discovery
The third area is software for lab orchestration, automated experiment design and systems that close the loop between hypotheses and results. Wintermute says founders building an autonomous layer for science are already selling into materials discovery and drug discovery labs.
It also describes quantum as a wildcard alongside that theme. Simulation and sensing could change what is discoverable in step-function fashion, while post-quantum security is already a real requirement at the settlement layer. The area is hard to insure and the winners are not obvious, the report says, but there is something there worth paying attention to.
R[3]sidency × Construct
Wintermute says the infrastructure required for the machine economy does not yet exist, and that is exactly where the work is. The firm says it wants to back founders who can see new problems emerging at the intersection of financial rails, autonomy and trust, and who can ship products on today’s rails while staying adaptable as standards evolve.
That is the pitch for R[3]sidency × Construct, which Wintermute describes as an accelerator for eight teams. Each team receives $300,000, spends 12 weeks in residence in London, gets access to more than 30 mentors, and presents at Demo Days in London and New York. The program is run with Fabric Ventures, Solana and Coinbase.
Wintermute closes by saying it wants to support founders building for a world where machines and humans trade and operate side by side.