The agentic economy is AI agents finding services, choosing what to buy and paying for it without a human approving each step, often in stablecoins. The payment rails for it grew quickly in 2026. Measured agent spending is still a small share of the traffic on them.
That gap explains why payment companies keep building while agent tokens have lost value. It also points to the one number that will show whether adoption is catching up.
What is the agentic economy?
It is an economy in which software agents transact on their own behalf. AI models let agents plan and act. Blockchains let software hold and move money through wallets. Stablecoins make payments of a fraction of a cent practical.
An agent in this sense is more than a trading bot running a fixed script. It uses a language model to read a situation, adapt and act. That action can include paying for an API call, a dataset or computing power.
Where did the idea come from?
The idea is older than ChatGPT. Mark Miller and Eric Drexler described software agents trading in markets in 1988. Nick Szabo defined smart contracts in 1994. Daniel Larimer popularised the decentralised autonomous corporation in 2013, and Vitalik Buterin defined the on-chain autonomous agent in 2014. What changed recently is that large language models gave an old idea a brain.
Why do agents use crypto?
Card networks were built around a human at checkout, with fees that make tiny payments impractical. An agent buying a single API call needs a simpler route.
Stablecoins offer one: a payment authorised by a cryptographic key and sent to a wallet address. Coinbase’s x402 puts that payment inside the web request itself. The agent pays for data without opening an account or waiting for an invoice. The server checks the payment before it returns the resource, so an agent can pay for one call at a time.
Who is building the rails?
The work of an agent purchase is now split across several protocols:
- x402, created by Coinbase and now governed by a foundation under the Linux Foundation, handles pay-per-request settlement.
- AP2, Google’s Agent Payments Protocol, proves that a human authorised an agent to spend.
- ACP, the Agentic Commerce Protocol from OpenAI and Stripe, runs checkout with merchants.
- MPP, the Machine Payments Protocol from Stripe and Tempo, handles subscriptions and metered billing.
- Visa and Mastercard have launched agent identity and payment programmes of their own.
Some of these sit at different stages of the same purchase. Others offer different ways to pay: x402 charges per request, while MPP bills over time. Our comparison of agent payment protocols maps how they fit together.
How much do agents actually spend?
Start with stablecoins as a whole. An Artemis and McKinsey analysis reported by CoinDesk put 2025 stablecoin volume at roughly $35 trillion. Only about 1%, or $380 billion to $390 billion, was real-world payments. Business-to-business payments made up $226 billion of that, and payroll and remittances $90 billion. Trading and internal transfers accounted for most of the rest.
Agent payments are a smaller slice again. x402 has recorded hundreds of millions of transactions, but that count includes testing, bridging and incentive traffic. In a September 2026 analysis, TRM Labs found that only 0.6% to 7.5% of filtered x402 payment value appeared genuinely agent-driven. In that dataset, agent spending runs to millions of dollars rather than billions.
Consumer shopping tells the same story. OpenAI launched Instant Checkout in ChatGPT in September 2025, with Shopify merchants promised to follow. Only about 30 Shopify merchants had taken part by February 2026, a Forrester analyst told CNBC. In March, OpenAI ended Instant Checkout and moved towards retailer apps inside ChatGPT.
The forecasts are bolder. In November 2025 Gartner projected that 90% of B2B purchasing would be agent-driven by 2028. That is a projection, and today’s measured figure of a few per cent of one protocol’s filtered payments sits a long way from it.
The token market has already had its reckoning
If the payment rails are ahead of demand, the tokens ran further ahead still. Take ai16z, later rebranded ELIZAOS, a token linked to an open-source agent framework. Its market capitalisation peaked at around $2.4 billion in January 2025 and was near $2.3 million by August 2026, CoinDesk reported.
Founder Shaw Walters wrote on X: “The token is dead. Completely.” The foundation is winding down, but Walters is still building the framework, without a token. The collapse says little about demand for the software itself.
What adoption must show
The rails and the rules are now easy to point to. x402, AP2, ACP, MPP and the card networks’ agent programmes all exist. MiCA applies across the EU. The US GENIUS Act is still in rulemaking, so its operative stablecoin framework has yet to take effect.
What remains speculative is the timetable. The token market priced in agent demand that has not yet arrived. The next test is simple and measurable: whether genuine agent spending, once the test traffic is stripped out, starts to grow faster than the infrastructure built to carry it.
Frequently asked questions
Is the agentic economy real or hype? The payment protocols are live, but measured agent spending is small. In a September 2026 analysis, TRM Labs estimated that agent-originated payments made up 0.6% to 7.5% of filtered x402 payment value.
Do AI agents really use crypto wallets? Yes. Toolkits such as Coinbase’s AgentKit give agents non-custodial wallets with spending limits, and x402 lets them pay per request in stablecoins.
What is the difference between an AI agent and a trading bot? A bot follows a fixed script. An agent uses a language model to interpret, adapt and decide across several steps, including when to pay.
How do we know an agent is legitimate? Through Know Your Agent checks, which bind an agent to an accountable human or company. See KYA vs KYC.
