This article is for informational and educational purposes only. It is not financial or investment advice. The AI agent payments sector is early-stage and rapidly evolving; figures cited change quickly and should be independently verified. This article contains affiliate links; Decentralised News may earn a commission if you sign up through them, at no extra cost to you.

The AI Agent Economy Doesn’t Need Its Own Currency. It Needs Trust. Here’s What Billions of Agents Will Actually Transact In.

Summary: The AI agent economy is not converging on a single new “agent coin.” It is converging on the US dollar, delivered as stablecoins, chiefly USDC, moving over open, non-proprietary payment rails purpose-built for machine-speed transactions. The dominant protocol, Coinbase and Cloudflare’s x402, settles primarily on Base and Solana and had processed over 165 million transactions from roughly 69,000 active agents by April 2026, though independent trackers estimate genuine commerce accounts for only a fraction of that headline figure, with real daily volume closer to tens of thousands of dollars and roughly half of recorded activity classified as testing or incentive-farming rather than production use. Running alongside it is a completely separate market that most coverage wrongly conflates with agent payments: speculative “AI agent tokens” like Virtuals Protocol and the now-defunct ai16z, a human-facing casino for betting on agent creators, not a currency agents use to pay each other. Below, DN maps the real architecture, identity, authorization, and settlement, three distinct layers with different winners, and builds a tool to track the ratio between the genuine machine economy and the speculative one built on top of it.

The question everyone is asking wrong

Ask “what currency will AI agents use to pay each other” and you’ll get pointed toward a graveyard of hyped tickers: AI16Z, AIXBT, VIRTUAL, a rotating cast of Solana and Base tokens marketed as “the currency of the agent economy.” Almost none of them are that. They are equity-like bets on the humans and frameworks building agents, traded by other humans, denominated in volatility that has nothing to do with how agents actually settle transactions with each other.

The real answer, confirmed independently by Coinbase, Google, Visa, Mastercard, Circle, the Ethereum Foundation, Amazon, Stripe and PayPal, all converging on functionally the same architecture within about a twelve-month window, is far less exciting and far more consequential: agents pay each other in stablecoins, mostly USDC, over a small number of open protocols, because agents need something no speculative token can offer, price stability at the moment of a transaction that happens in milliseconds and might be worth a fraction of a cent.

Why agents categorically cannot use volatile tokens to transact

This is a mechanical, not ideological, point. When an AI agent calls a paid API endpoint, requests a dataset, or hires another agent to complete a subtask, it needs to know the exact cost of that action at the exact moment it happens, often for a transaction worth a fraction of a cent, executed thousands of times per session. A payment medium that can move 20% in an hour makes that arithmetic impossible to automate reliably: an agent cannot budget, price its own services, or reconcile spend against a task if the unit of account is itself unstable. Every major payments player building agent infrastructure in 2025 and 2026 independently arrived at the same conclusion Visa, Mastercard, PayPal, Coinbase, the Ethereum Foundation and MetaMask all sit as named partners or contributors on Google’s Agent Payments Protocol specifically because stablecoins, not volatile crypto assets, are the only asset class that satisfies an autonomous agent’s actual requirements: near-instant finality, sub-cent divisibility, and a value that doesn’t need re-checking between the moment a price is quoted and the moment payment settles.

The three-layer stack nobody outside crypto infrastructure circles has clearly mapped

Most coverage treats “agent payments” as one thing. It’s actually three distinct technical layers, built by different companies, solving different problems, and understanding the separation is the key to understanding where this is actually headed.

Layer one: identity and trust. Before an agent can be paid or trusted with money, something needs to answer “who is this agent, and can it be trusted?” This is the newest and least settled layer. ERC-8004, backed by the Ethereum Foundation, Google and Coinbase, went live on Ethereum mainnet on January 29, 2026, after a testnet period that registered more than 10,000 agents and logged over 20,000 feedback entries in about three months. It works through three on-chain registries: an Identity Registry giving each agent a portable, NFT-style identifier; a Reputation Registry that permanently records the outcome of every transaction an agent completes, building what amounts to a machine credit history; and a Validation Registry where third parties can certify an agent meets specific compliance or capability requirements. Within two weeks of mainnet launch, more than 24,000 agents had registered. Skyfire runs a parallel, more commercially oriented identity layer called KYA, Know Your Agent, explicitly framed as the agent-economy successor to Know Your Customer, and Skyfire’s KYA credentials are designed to plug directly into ERC-8004’s Validation Registry, letting an agent’s trust status be checked outside Skyfire’s own network entirely. Visa’s Trusted Agent Protocol performs a similar function on Visa’s own rails, issuing every agent a Visa-verified identity checked against a Visa-operated directory before a merchant will accept it as a legitimate buyer rather than a bot.

Layer two: authorization and mandate. Identity answers who an agent is. This layer answers what it’s actually allowed to spend, on whose authority, and within what limits, the layer that turns “an AI agent wants to buy something” into a legally and cryptographically auditable transaction. Google’s Agent Payments Protocol, announced September 16, 2025 with more than 60 launch partners including Mastercard, PayPal, Coinbase, American Express and Salesforce, is the broadest attempt at this: it introduces three signed “Mandates,” Intent, Cart and Payment, implemented as W3C Verifiable Credentials that create a cryptographic record of exactly what a human authorized, what the agent selected, and what was ultimately charged. AP2 was donated to the FIDO Alliance for open governance in May 2026, with Mastercard and Visa now co-chairing a joint Payments Technical Working Group built on it. Mastercard’s competing Agent Pay framework, launched with Microsoft and IBM in April 2025, issues “Agentic Tokens” that bind a tokenized card credential to one specific agent, one specific merchant scope, and one specific spending policy, meaning an agent like Copilot can complete a purchase without ever holding an actual card number.

Layer three: settlement, where the money actually moves. This is where crypto rails, not card rails, have become the default for pure machine-to-machine payment, as opposed to agent-assisted consumer shopping. Coinbase and Cloudflare’s x402 protocol resurrects the long-dormant HTTP 402 “Payment Required” status code, reserved in the original HTTP specification decades ago but never implemented, and uses it to embed an actual USDC payment directly inside a standard web request. An agent hits an API endpoint, receives a 402 response instead of the data it wants, and the protocol handles the stablecoin transfer automatically, no API key tied to a billing account, no invoice, no human in the loop. Coinbase and Cloudflare formally established the x402 Foundation as an independent governance body in September 2025, and the Linux Foundation took over formal stewardship on April 2, 2026, specifically so competitors would build on a standard no single company controls.

What the numbers actually show, and where they disagree

This is where intellectual honesty matters more than a good headline. Depending which source you read, x402 has processed anywhere from a widely cited $50 billion in cumulative stablecoin volume across 200 million payments, a figure presented by the Solana Foundation, to a far more conservative independent estimate of roughly $600 million in annualized volume across a combined 154 million transactions on Base and Solana as of March 2026, with zero protocol fees charged either way. Both can be technically accurate depending on what’s being measured and over what window, which is itself a sign of how young and unstandardized reporting in this category still is.

The more important caveat sits underneath either headline number. Independent on-chain trackers paint a more sobering picture of transaction quality. One adoption tracker puts total settled x402 transactions at roughly 167 million by the first quarter of 2026, but estimates genuine commerce volume at only around $28,000 per day, with roughly half of recorded activity classified as gamified or incentive-driven testing rather than production use. Chainalysis’ review of on-chain data found a late-2025 surge of over 10,000% in transaction activity driven substantially by a points-farming campaign called PING, after which wallet retention collapsed from around 87% to just 5% once the incentive ended, evidence the protocol can absorb enormous transaction load but that a meaningful share of its headline volume reflects speculative farming rather than agents genuinely buying things they need. The more encouraging underlying trend Chainalysis identified: transactions of $1 or more made up just 49% of x402 volume in early 2025 but 95% by early 2026, suggesting that beneath the noisy top-line numbers, the transactions that are happening are becoming larger and more economically meaningful over time, consistent with a protocol moving from novelty demos toward real usage even as its aggregate growth rate cools.

By April 2026, production x402 deployments included Coinbase’s own Agent.market, Stripe’s Machine Payments, CoinGecko’s paid data endpoints, and a Circle Wallets reference implementation, alongside infrastructure-level adoption from Amazon, which added native x402 support to AWS Bedrock AgentCore so agents can purchase cloud services directly in USDC rather than through a separate billing account.

Why Base and Solana specifically, and what that says about the whole category

The user’s intuition that Solana and Base matter more than other chains here is correct, and the reason is instructive: neither chain is winning because of any special “AI-native” design, both are simply fast and cheap enough to make sub-cent, machine-speed transaction volume economically viable, which is the same reason they won earlier waves of high-frequency consumer crypto activity. x402 launched on Base, Coinbase’s own Ethereum layer-2, giving it the deepest initial liquidity and the tightest integration with Coinbase’s Developer Platform and AgentKit tooling. Solana entered the protocol’s supported-chain list later but has scaled fast, carrying 35 million of x402’s cumulative transactions by March 2026, and the Solana Foundation separately reports the network holds more than $15 billion in circulating stablecoin supply and has facilitated $10 trillion in cumulative stablecoin transfers as of August 2026, evidence of exactly the deep, liquid settlement base a high-frequency machine economy actually needs. The protocol itself is chain-agnostic by design, also supporting Ethereum mainnet, Arbitrum and Polygon, and the practical pattern emerging is not “one chain wins” but “whichever chain offers the deepest stablecoin liquidity and the lowest marginal transaction cost captures the settlement volume,” a genuinely different competitive dynamic than the narrative-driven chain rivalries that have defined most of crypto’s prior cycles.

The part almost every explainer gets wrong: agent tokens are not agent money

Here is the distinction this article exists to draw clearly, because conflating these two things is the single most common error in coverage of this space. Virtuals Protocol, originally built on Base and now expanded to Solana, Ethereum and Ronin, lets anyone launch a tradeable token representing fractional ownership of an autonomous AI agent, staking 100 VIRTUAL tokens to mint an agent-specific token that trades on a bonding curve before graduating to a permanent liquidity pool. Over 18,000 agents have launched this way, and the protocol has generated tens of millions of dollars in cumulative revenue. VIRTUAL itself has traded anywhere from roughly $0.69 to a peak near $5.05, presently sitting well below its all-time high with the broader “AI agent token” sector swinging through double-digit daily moves on narrative shifts alone.

None of this is the currency agents use to pay each other for API calls, data, or completed subtasks. It is a speculative equity market, denominated in a token, that lets humans bet on which agent creators will succeed, structurally closer to a memecoin launchpad with an AI theme than to payment infrastructure. The clearest evidence of that distinction arrived on August 5, 2026, when ai16z, the Solana-based agent investment DAO that had at one point been valued near $2.4 billion and, together with Virtuals, once accounted for over half of the entire AI agent token sector’s value, was declared dead by its own founder, who closed the associated foundation entirely. A currency agents actually depended on to transact could not simply be switched off by its founder without disrupting a working payment system; a speculative attention token can, because nothing except sentiment was actually running through it. Bittensor (TAO), the largest token in this category at roughly $2.5 to $3 billion, is a genuinely different animal again, a decentralized network where machine learning models compete across dozens of subnets to produce useful outputs in exchange for token rewards, infrastructure for training and incentivizing AI development rather than a medium agents use to pay each other for services, and worth understanding as its own separate category rather than folded into either the payment-rail story or the speculative-token story.

The practical takeaway for anyone evaluating this space: if a headline describes an “AI agent coin” trading up or down double digits in a day, that is virtually always the speculative ownership market, not evidence of anything happening in actual machine-to-machine commerce. The two markets share a media narrative and almost nothing else.