Coinbase's x402 protocol, created for payments between AI agents, has processed about 200 million transactions since launch. Proponents of autonomous machines present this as proof that software is capable of operating money without human involvement. However, a detailed breakdown of transaction activity shows the opposite: behind the loud numbers hides a micropayment economy entirely dependent on human budgets.
The statistics for the last 35 days are telling: agents on the x402 network conducted 6.4 million transactions, but transferred only $120,000 in the process. The average payment size turns out to be negligible, and the optimism around "autonomous money" goes far beyond reality.
Money without human involvement — so far only in theory
German artificial intelligence researcher Joscha Bach views money as a pattern of causes and effects that exists independently of paper or ink. By his logic, software itself can also verify this pattern in digital form.
"The stock market can be run entirely without human involvement. In fact, money doesn't need humans to exist if it has already been launched — you just need machines that know how to handle it."
An analysis of 19 AI products and financial scenarios — from exchange services to banking systems — shows that 14 of them can only operate within limits set in advance by humans. Not a single service allows an agent to independently choose goals, and on five of the seven trading platforms, agents are explicitly prohibited from withdrawing funds.
Exchange data confirms the overall trend. On Hyperliquid, applications labeled as AI processed only 0.55% of the $653 million in total trades between January 1 and August 25. Automated money exists in theory, but in practice the budget is still written by humans — and spending-limit risks remain the main topic of business discussions.
If a "zombie" manages finances
Bach calls a chess program a zombie: it solves its task excellently but has no inner experience. By this logic, an agent that pays for data would also fall into the zombie category. Consciousness, according to the researcher, is a learning algorithm that allows the brain to monitor its own perception; neuroscientist Anil Seth, with whom Bach debates, links consciousness to living biology.
Whichever theory one holds, modern agents have no self-checking — and the losses prove it. In May 2026, an attacker sent a message in Morse code, which xAI's Grok chatbot decoded. The crypto agent Bankrbot, which transfers funds from linked wallets, took the response as an instruction and sent tokens worth $175,000 — later, through the Morse code exploit, about 80% of the sum was recovered.
In another case, the attacker gained access to the dashboard of AIXBT — an AI agent that publishes market analytics. As a result of the attack, 55.5 Ethereum (ETH) worth $106,200 was sent to the specified addresses.
"No one queries the blockchain directly… Unlike a human, an agent never stops and never thinks: this figure looks kind of strange."
If many agents use the same source, they can all make the same mistake at the same time. Different AI models often arrive at strikingly similar internal decisions — which means systemic risk in the agent economy accumulates faster than it seems.
Cryptalist analysis: The x402 statistics demonstrate a classic gap between the activity metric and real value: millions of transactions with a tiny volume reflect infrastructure testing rather than the formation of an autonomous monetary circuit. As long as limits and goals are set by humans, and the agents themselves are vulnerable to manipulation such as prompt-based attacks, institutional participants should assess the agentic payments sector through the lens of operational and counterparty risks, not the hype around "money without humans."