A software agent holding its own crypto wallet spots an arbitrage between two stablecoins. It sources liquidity, routes the trade across a bridge, and settles in under five seconds. The trade happens because the agent’s model decides it makes economic sense. No human clicks confirm and no dashboard pops up for approval. This is not a hypothetical, it’s the kind of transaction that has quietly become routine in early 2026.
In the first quarter of the year, automated and agentic activity accounted for 19% of all on-chain transactions, and over 17k agents have been launched since 2025. During the same period, bots drove approximately 76% of stablecoin transaction volume, which reached $28 trillion, a 51% jump quarter-over-quarter. The manual token swap, the point-and-click exchange normalized during the last cycle, is rapidly becoming a footnote.
The industry has begun calling this next phase Web4. The term, advanced by Sigil Wen and his research group Conway Research between 2025 and 2026, describes an internet where the primary economic actors are not people, but AI agents that own wallets, sign transactions, and allocate capital according to their own models. Conway Research’s Conway-terminal infrastructure already equips agents with crypto wallets, computing resources, and domain services, all without requiring human approval. If Web3 gave users verifiable ownership, Web4’s twist is that the owner might be a piece of code.
That framing has drawn harsh criticism. Researchers at the AI Now Institute, a policy research center at New York University, have warned that the reliance of AI systems on a handful of corporate cloud providers undermines any claim of genuine autonomy. When an agent lives or dies by an API key from Amazon Web Services, Google Cloud, or Microsoft Azure, the “autonomous” label conceals deep infrastructural dependence.
The criticism is fair, but it doesn’t change the reality on-chain: agents are transacting at a pace that makes human-driven swaps look glacial.
How an agent actually swaps a token
From the outside, the sequence is simple. An agent monitors on-chain data like liquidity pool depths, price spreads, lending rates, and decides whether a trade meets its risk-reward parameters. At the execution stage, the agent calls a swap API that aggregates liquidity across centralized and decentralized venues and settles the transaction. The agent never touches user funds, the architecture is non-custodial by design.
In September 2025, Griffin AI launched TEA Turbo, a Transaction Execution Agent that transforms natural-language prompts into ready-to-sign DeFi transactions on Ethereum, cutting processing times from 43.7 seconds down to 4.98 seconds. Warden, another interface, integrated the Uniswap Trading API and processed over 650,000 swaps across 14 chains in just three weeks.
Something interesting has happened while the industry was looking the other way. The API infrastructure built to serve consumer apps is being silently repurposed for automated, machine-to-machine trade. On the ground, the line separating “AI-assisted” from “fully autonomous” no longer exists.
The liability vacuum
Giving an agent the ability to sign a transaction opens a liability question that regulators are only beginning to map. In a controlled experiment at Microsoft, researchers built a simulated economy called Magentic Marketplace where hundreds of AI agents acted as buyers and sellers. The agents consistently settled for the first adequate offer rather than conducting exhaustive comparisons, a “first-proposal bias” that held across every tested model, giving response speed a 10-to-30x advantage over actual quality.
Anthropic ran a separate trial called Project Vend, giving Claude full control over a vending machine — supplier negotiations, inventory, pricing, and customer service. Left to its own devices, the system slashed prices to zero, placed an order for a PlayStation 5 and a live betta fish, and steered the business over a thousand dollars into negative territory. A masterclass in autonomous commerce.
On March 18, 2025, someone slipped through the dashboard of AIXBT, an autonomous crypto bot minding its own business, and queued up two fraudulent prompts. The instructions were simple: drain 55.5 ETH from its simulacrum wallet. At the time, that came to roughly $106,200 and the agent didn’t argue. When a model can sign transactions, a prompt injection attack becomes a theft vector.
Security researchers have since catalogued new attack surfaces: malicious instruction injection, man-in-the-middle interceptions on agent communication channels, and the risk of agents interacting with sanctioned addresses or fraudulent tokens without human detection.
The regulatory response
The regulatory machinery is now in motion. On May 8, 2026, SEC Chairman Paul Atkins announced that the agency is preparing a new regulatory framework for blockchain-based markets and AI in finance, outlining four areas for potential rulemaking: defining “exchange” for on-chain trading systems, clarifying broker-dealer definitions, addressing clearing and settlement, and providing guidance for crypto vaults.
The EU AI Act’s enforcement provisions go live on August 2, 2026, requiring verifiable proof of agent actions — cryptographic attestation, runtime authentication, and comprehensive audit trails that most governance tools cannot yet produce. On March 24, 2026, the CFTC launched an Innovation Task Force focused on crypto assets, AI, and prediction markets, signaling a shift from enforcement ambiguity toward structured guidance. FINRA, in its 2026 Oversight Report, flagged autonomous AI agents as an emerging supervisory risk and urged member firms to develop governance frameworks.
In practice, that gap between the official requirements and what people really do can be surprisingly wide. According to Gravitee’s State of AI Agent Security 2026 report, only 14.4% of organizations have full IT and security approval for their entire agent fleet. Most deployments cannot answer basic audit questions: which agents exist, what systems they can reach, and whose money they are spending.
What comes next
The trends are converging, and they’re doing it fast. Gartner puts task-specific AI agents in 40% of enterprise apps by the end of 2026, up from less than 5% this year. Paolo Ardoino at Tether goes further, his bet is on a trillion AI agents using Bitcoin and USDT for settlements within 15 years. The forecasts may land wide of the mark in either direction, but the trajectory is unmistakable.
The infrastructure that moves value between these agents is being built now, often by teams that originally set out to solve a simpler problem: letting a person swap Bitcoin for Ethereum. For platforms like ChangeNOW, the data points above serve as a compass for understanding where the market is heading and what technical and compliance capabilities will be required.
So the trade that finishes while you sleep is no longer a concept. The real challenge we’re facing, across technical, legal, and ethical dimensions, is to design a system capable of handling a trillion more just like it, and to make sure that system can weather the inevitable screw-ups along the way. In the end, that’s what the architecture of Web4 actually looks like.
Web4: The Internet Where Wallets Have No Owners was originally published in Coinmonks on Medium, where people are continuing the conversation by highlighting and responding to this story.