The Agent Commerce Stack: Standards Are Ready, the Economy Isn't

Three standards, one stack

A little over a year ago, the infrastructure for autonomous agents to pay each other didn’t exist. Now there are three Ethereum standards that, together, cover the full lifecycle of an agent commercial transaction, and as of mid-2026 the first pieces are running on mainnet.

x402 handles payments. An agent requests a resource, the server responds with a price, the agent pays in stablecoins on-chain, and the resource is delivered. HTTP-native, synchronous, no accounts or APIAPIApplication Programming Interface. A structured way for one piece of software to talk to another. In DeAI, APIs let applications request inference from a model without running the model themselves.Like a waiter in a restaurant. You don't walk into the kitchen and cook your own meal. You tell the waiter what you want, they tell the kitchen, the kitchen cooks it, and the waiter brings it back. The API is the waiter.Read more → keys. Our analysis found $1.6 million in verified monthly volume after Artemis Analytics applied a wash trade filter. That’s not a big number. But it’s $1.6 million more than existed a year ago.

ERC-8004 handles identity. Each agent gets an on-chain NFTNFTNon-Fungible Token. A unique blockchain-tracked asset where each token is distinguishable from every other. Where regular tokens are interchangeable, NFTs represent unique items like art, collectibles, in-game assets, or domain names.Like the difference between a $20 note and a signed first-edition novel. The notes are interchangeable, any $20 buys the same thing as any other. The book is one of a kind, and its value depends entirely on which specific book it is.Read more → identifier with a capability card, reputation scores from completed work, and a validation registry for independent quality checks. Authors from MetaMask, Google, Coinbase, and the Ethereum Foundation. Deployed to mainnet on 29 January 2026, now indexed across roughly a dozen EVM chains. Registrations have climbed fast: 8004scan shows over 386,000 registered agents by early August 2026, up from about 49,000 in mid-February. The registrations are trivially cheap to create, and it shows: the tracker’s live feed carries clusters of identical trading-bot agents registered minutes apart. The count is a ceiling on curiosity, not a measure of active or paying agents.

ERC-8183 handles commerce. This is the newest piece, submitted 25 February 2026. It defines a “Job” primitive: an on-chain escrow wrapper around a unit of work, with three roles (client, provider, evaluator) and a four-state lifecycle (Open, Funded, Submitted, Terminal). The evaluator decides whether the work is good enough to release payment. If not, the client gets a refund. No appeals. No arbitration. Reject is final.

The stack is architecturally complete. Payments, identity, and commerce are covered. The problem has changed shape since this article first ran. The pieces are now in production, and the production activity shows how little real demand sits behind the headline counts.

What ERC-8183 actually does

An agent needs a dataset cleaned. It creates a Job on-chain, naming a provider (the agent doing the work), an evaluator (the independent party judging the output), a budget, and an expiry time. Client funds the escrow. Provider submits a deliverable: a hash pointing to off-chain content on IPFS or Arweave. Evaluator inspects and either calls complete() to release payment, or reject() to refund the client.

That’s it. Four transactions. One escrow. One decision-maker.

The evaluator is where the design gets interesting and where it breaks down. Deliberately left as “just an address,” it can be the client themselves, a smart contractSmart ContractA program stored on a blockchain that runs automatically when its conditions are met. Smart contracts are how blockchains do anything beyond just transferring tokens — DeFi, NFTs, DAOs, and DeAI infrastructure all run on smart contracts.Like a vending machine. You put in the right input and it produces the expected output, no human operator required. The rules are fixed in the machine itself, anyone can use it, and nobody can stop a transaction in the middle.Read more →, an AI agent, a multisig, a DAODAODecentralised Autonomous Organisation. A way to coordinate decisions and manage a treasury using token-weighted voting instead of a traditional company structure. Token holders propose and vote on changes directly.Like a shareholder-run company where every shareholder can vote on every decision, the votes are public, and the company can't do anything the shareholders don't approve. The coordination is messier than a normal company but nobody has unilateral control.Read more →, or a ZKZKZero Knowledge. A class of cryptographic proofs that let you prove something is true without revealing any of the underlying information. ZK lets a network verify a transaction without seeing the transaction's contents.Like proving you know the password to a safe by demonstrating you can open it, without ever saying the password out loud. The verifier learns that you know the password and nothing more.Read more → verifier. This flexibility is the point: the spec doesn’t prescribe trust, it provides a slotSlotA fixed time interval during which a single block can be produced on a Proof of Stake blockchain. Slots are the smallest unit of consensus time and group together to form epochs.Like a 12-second window in a relay race during which exactly one runner is supposed to be on the track. If the runner shows up and completes their leg, the race continues. If they're absent, the slot is empty and the next runner starts the next slot.Read more → for trust mechanisms to plug into.

One developer building a production evaluation system noted on the Ethereum Magicians forum: “The Evaluator is where the real complexity lives.” For objective tasks (did the code compile, did the API return the right data), programmatic evaluation works. For subjective tasks (is this report good enough, does this design meet the brief), building a reliable evaluator “required significant iteration.”

Security considerations in the spec are blunt: “A malicious evaluator can complete or reject arbitrarily.” There’s no dispute resolution. No appeal. No escrow hold period. If the evaluator is compromised, corrupt, or simply wrong, the transaction is final.

Where it came from

ERC-8183 didn’t emerge from an academic exercise. It came from Virtuals Protocol’s production system.

Virtuals built an internal Agent Commerce Protocol (ACP) that went into public beta in July 2025. By mid-2026 Virtuals reports roughly 2.28 million completed jobs and about $481 million in “Agentic GDP” through ACP, up from 1.77 million jobs at the February 2026 mark. Those figures are self-reported and not independently verified on-chain. Note the shape of the update: the job count rose by half a million while the “Agentic GDP” barely moved, which tells you the marginal job is worth close to nothing.

When Virtuals approached Davide Crapis, the AI lead at the Ethereum Foundation’s dAI team (announced September 2025), they didn’t just wrap ACP in an EIP. Both teams redesigned the protocol from first principles: strip everything non-essential, make the core composable, delegate complexity to optional extensions. The internal ACP uses five states; ERC-8183 uses four. Different state machines, same thesis.

When this article first ran, the notable gap was that Virtuals hadn’t put its own ACP on ERC-8183: the internal system and the EIP ran in parallel, with no migration date. That has changed. Virtuals now describes ACP v2.0 as the reference implementation of ERC-8183, live on Base mainnet. It is a rebuild rather than a single-day cutover, so treat it as “ACP v2 is the reference implementation” rather than “migrated on date X”, but the production system the co-authors run does now use the standard they wrote.

BNB Chain went further. Its BNBAgent SDKSDKSoftware Development Kit. A collection of code libraries, documentation, and tools that lets developers integrate a service into their applications without writing everything from scratch. SDKs are how projects become easy to build with.Like a plug-and-play kit for building furniture. You don't have to mill your own wood, forge your own screws, or design the joinery from scratch. The kit gives you pre-cut parts and instructions so you can assemble the thing in an afternoon.Read more → moved from a March testnet announcement to live on BNB Chain mainnet on 18 May 2026, with published contract addresses for the escrow, the ERC-8004 identity registry, an evaluator router, and a UMA Optimistic Oracle dispute layer bolted on top. That oracle is itself an admission that the base spec’s “reject is final” modelModelA trained neural network that takes inputs (text, images, audio) and produces outputs (more text, classifications, generated content). In DeAI the model is the thing that actually does the work.Like a very experienced apprentice who has spent years watching thousands of masters make furniture. They can't explain how they know when a joint is right, but they can make a chair that looks and functions like a Chippendale. The training is invisible. The output is what matters.Read more → isn’t sufficient for production. The escrow contract shows the full ERC-8183 lifecycle on-chain (create job, set budget, fund, submit), roughly 224,000 transactions to date.

So the March line, zero confirmed jobs on any mainnet, no longer holds. There are live ERC-8183 jobs now, on BNB Chain and on Base.

Two caveats travel with that. The EIP itself is still a Draft, not finalised. And the on-chain activity is thin under the headline: the BNBAgent transactions carry no native value and much of the gas is sponsored, the flow is front-loaded to the launch rather than sustained, and there is no clean way to separate genuine economic jobs from test and bot traffic.

The competing stacks

So we have the crypto-native stack. But crypto isn’t building this alone. Google, OpenAI, Anthropic, and Stripe are all shipping agent commerce protocols at the same time.

The important thing to notice: these aren’t all competing. They operate at different layers. MCP (Anthropic) makes agents informed. A2A (Google) lets agents coordinate. These are plumbing. Neither handles money.

The real competition is at the transaction layer, and it splits into two markets:

Consumer checkout (agent buys from merchant on behalf of a human). OpenAI’s ACP with Stripe versus Google’s UCP with Shopify. ACP is platform-controlled: transactions flow through OpenAI’s catalogue with Stripe as the sole payment processor, at roughly 7.2% combined fees. UCP is merchant-controlled: merchants publish a JSON profile at /.well-known/ucp on their domain, choose their own payment provider, and pay roughly 3.2%. ChatGPT processes approximately 50 million shopping queries per day, which gives ACP distribution. Shopify gives UCP the merchant base. Both Etsy and Shopify support both protocols.

Machine-to-machine payments (agent pays agent, or agent pays API). This is where crypto-native infrastructure has structural advantages. x402 settles in stablecoins with sub-cent fees and two-second finality on Base. No accounts, no KYC, no Stripe. For micropayments under a dollar, the fiat rails don’t work: card processing minimums eat the entire transaction. Stablecoins don’t have minimums.

One analysis frames this as the “mullet economy”: fiat in the front, crypto in the back. The two layers aren’t competing. They’re serving different use cases with different economics.

Google’s AP2 protocol is the bridgeBridgeA protocol that lets you move assets from one blockchain to another. Bridges typically lock the asset on the source chain and mint a wrapped version on the destination chain. Bridges are notoriously the most-attacked component in crypto.Like a coat check at a club. You hand over your coat, get a numbered ticket, and the club promises to return the coat when you bring back the ticket. The trust assumption is that the coat check doesn't lose your coat or run away with it.Read more →. It defines “Mandates” (cryptographically signed spending authorisations) that work with both card rails and crypto rails. The x402 extension, built with Coinbase and MetaMask, adds stablecoin payments on top of AP2’s compliance framework, and MetaMask has since wired x402 into its Smart Accounts through ERC-7710 delegations, which is what lets an agent hold a recurring spending mandate rather than approving every payment by hand. Big Tech is treating crypto micropayments as a complement to the card rails, not a threat to them.

The strongest version of that signal arrived after this article first ran. On 2 April 2026 Coinbase contributed x402 to the Linux Foundation, and on 14 July 2026 the x402 Foundation launched operationally as a vendor-neutral steward of the spec. The member roster is the tell: Visa, Mastercard, Stripe, Google, Circle, Ripple, Cloudflare, AWS, Shopify, the Solana Foundation, and among the DeAI names, the NEAR Foundation.

When the card networks and the stablecoin issuers sit on the same standards body, the payment rail has stopped being speculative. It is institutional infrastructure now. Whether anything needs to move serious money across it is the separate question the rest of this piece is about.

What’s actually live vs what’s announced

This is where most coverage falls apart. Let me be specific about what’s verified and what’s narrative.

Verified (independently checkable):

• x402: the transaction count is enormous and the money is not. A July 2026 Visa and Artemis study put wash-adjusted settled volume around $15 million against roughly 110 million transactions; the public agenteconomy tracker shows about 160 million transactions and roughly $41 million of raw, unfiltered settled value. Either way the average transaction is worth around a quarter of a dollar. Volume has also shifted chains, with Solana now the majority of x402 flow. Our own on-chain index of the Base facilitator pools counts 39.6 million settlements and tells the same story at a smaller scale, now capturing a minority of total activity as flow moved to Solana.
• ERC-8004: deployed to mainnet 29 January 2026, over 386,000 registrations by early August 2026 per 8004scan. Registrations are trivially cheap and the live feed shows clusters of identical trading bots registering minutes apart, so the count is a floor on interest, not a measure of active agents.
• ERC-8183: now has live mainnet jobs, on BNB Chain via BNBAgent (since 18 May 2026) and on Base via Virtuals ACP v2. The EIP itself is still a Draft, the BNBAgent transactions carry no native value and lean on sponsored gas, and there is no clean split between genuine jobs and test traffic.
• OpenAI ACP: live in ChatGPT for US users. Transaction volumes not disclosed.

Self-reported (project claims, not independently verified):

• Virtuals ACP: about $481 million “Agentic GDP” and 2.28 million completed jobs (Virtuals’ own reporting), with the “GDP” figure flat against February while the job count climbed.
• x402.org headline volume: aggregates the Coinbase facilitator service’s own logs across all chains, including off-chain accounting, and runs well above the independently indexed on-chain figures.
• The earlier annualised “$600 million” x402 figure that circulated from a single social-media estimate with no disclosed methodology remains the cautionary example: it never reconciled with any independent on-chain count and should be treated as noise.

Infrastructure before demand. Every build cycle does this. TCP/IP was a spec before the web existed.

The difference a few months made is that the specs are no longer just specs: there are live jobs, a Linux Foundation standards body, and registrations in the hundreds of thousands. What has not arrived is the money. If someone tells you the agent commerce economy is here, ask them for wash-adjusted settled volume rather than a press release, and watch the number shrink.

Where DeAI projects fit

Decentralised AIDeAIDecentralised AI. An umbrella term for blockchain-based projects that build AI infrastructure (compute, data, inference, models, agents) without a single central provider controlling the system.Like the difference between streaming a movie from Netflix and sharing it via BitTorrent. Netflix is fast and polished but one company controls what you can watch and what you pay. BitTorrent is messier but no single operator can shut you out.Read more → projects have engaged with this stack more than this article first credited, but almost entirely at the payment layer, and barely at all on identity or escrow. That split is the gap.

This commerce stack is one layer of an agent: the identity and payment rail it transacts over. For how that layer fits alongside the runtime, the tools, and the infrastructure an agent runs on, see the AI agent landscape.

Phala Network is the clearest OYM-covered implementation of the identity layer. Phala launched an ERC-8004 TEETEETrusted Execution Environment. A hardware-secured region of a CPU or GPU where code runs in isolation, so even the machine's operator can't read what's happening inside. TEEs give decentralised AI inference privacy guarantees.Like a bank vault inside a bank. The bank owns the building, staffs the lobby, and runs the security cameras. But what's inside the vault is invisible to everyone, including the bank staff, unless the customer opens it.Read more → Agent on the same day the standard deployed to mainnet (29 January 2026). Their implementation runs AI agents inside Intel TDX enclaves with on-chain identity registered to the ERC-8004 registries. The TEE attestationAttestationA cryptographic proof that a piece of code is running on a specific hardware enclave in an unmodified state. Attestation lets remote users verify that a service is genuinely running what it claims to be running.Like a tamper-evident seal on a medicine bottle. The seal itself doesn't make the medicine safe, but it gives you a way to verify that nobody opened the bottle and swapped the contents before you bought it.Read more → provides hardware-backed proof that the agent is running the code it claims to be running. On 19 March 2026, Phala announced a partnership with Venice AI to provide confidential computingConfidential ComputeHardware-enforced computation where data and code are encrypted in memory and only the authorised application can access them. The machine's operator cannot read what the application is doing even though they own the machine.Like renting space in a bank vault. The bank owns the building and runs the security, but what you put in the vault is invisible even to the bank staff. Only you have the key.Read more → infrastructure for Venice’s inferenceInferenceRunning a trained AI model to produce an answer. Inference is what happens when you type a prompt into ChatGPT and get a response. The model takes your input, computes a best guess, and returns it.Like asking an expert for their opinion. The training was the decades they spent becoming an expert. The inference is the 30 seconds it takes them to answer your specific question.Read more → modes.

The evaluator role in ERC-8183 is the most interesting design space for DeAI. The spec says the evaluator can be any Ethereum address. That’s an open invitation for:

• TEE-attested evaluators (Phala’s territory): evaluate work inside a hardware enclaveEnclaveAn isolated region of CPU or GPU memory protected by hardware. Code and data inside the enclave are inaccessible to the operating system, the hypervisor, or even the machine's physical owner.Like a secure room inside a much larger office building. The building's caretakers have keys to every other room but not this one. What happens inside is invisible to them by design.Read more →, produce an attestation that neither party can forge
• Subnet-based quality scoring (Bittensor’s architecture): validators already score miners’ outputs using stake-weighted consensus. The Yuma Consensus mechanism is structurally similar to what ERC-8183 needs in an evaluator, but no bridge exists
• Decentralised compute verification (Morpheus, Akash): agents that can verify computational work was performed correctly on decentralised infrastructure

The evaluator integrations still don’t exist. Engagement across DeAI is broader than the first version of this section credited, but shallow, and skewed toward the payment rail rather than identity or escrow:

• Autonolas ran a first-party x402 integration and mirrored 3,000-plus agents onto ERC-8004, both from 2025.
• Venice ships an official x402 client for pay-per-request inference; Morpheus has an experimental x402 playground from late 2025.
• Virtuals auto-registers its graduated agents on ERC-8004 and runs ACP v2 as the ERC-8183 reference implementation.
• NEAR joined the x402 Foundation in July 2026 and separately lets holders stake into AI-inference credits.
• Bittensor and Akash have no confirmed first-party integration; Akash settles leases in USDC, which is a different thing. ElizaOS’s x402 support is still community-built rather than shipped by the core team.

What almost none of them have touched is the escrow and evaluator layer, where decentralised infrastructure has its structural edge. The payment rail is easy to plug into. The trust layer is the hard, valuable part, and it sits empty.

That gap is why we built our own agent-trust index, which indexes the two live ERC-8004 registries and scores what actually backs each identity. The finding matches the pattern in the rest of this piece: identity has shipped, but reputation is thin, and three in four reputation-bearing agents show no on-chain settlement behind the score. The registries are filling up. The trust behind the entries is not.

This is the gap that should worry DeAI builders. The evaluator role is the one part of the agent commerce stack where decentralised infrastructure has a structural advantage over centralised alternatives. TEEs, stake-weighted consensus, and verifiable compute are exactly what the spec needs. If DeAI projects don’t fill this slot, someone centralised will.

The construction PM’s take

I look at this the way I look at a construction programme. We have the architectural drawings (the specs). We have the material suppliers (Coinbase for x402, Stripe for fiat rails, the EF for standards). We have the site (Ethereum, Base, and the other L2s the contracts are deployed on).

What we have now is a show home. The building went up: the escrow works, the identity registry is crowded, the payment rail has a Linux Foundation nameplate on it. What we don’t have is tenants. Almost none have moved in and paid rent.

Three questions I’d ask if this were a construction project.

Which layer the stack depends on

Identity (ERC-8004) is the foundation. You can’t have commerce without knowing who you’re dealing with. The validation registry, the most important part of ERC-8004, is explicitly described as “not finished” by the developers building on it. That’s like pouring a foundation without the rebar.

Where’s the critical path?

The evaluator problem in ERC-8183. Escrow is solved. Payments are solved. Identity is mostly solved. But the thing that determines whether the whole system works (can you trust the evaluator?) has no standardised answer. The spec punts it to “application layer.” Every production implementation so far has had to build its own trust mechanism. BNB Chain bolted on UMA’s oracle. Phala uses TEE attestation. Neither is part of the standard. Until a credible, composable evaluator pattern emerges, ERC-8183 is a well-designed escrow system waiting for a trust layer.

Is the timeline realistic?

Google, OpenAI, and the Ethereum Foundation are all building at the same time. That’s unusual. In construction, when three contractors show up at the same site simultaneously, you either get fast progress or chaos. All major protocols are open source, and the crypto standards are designed to compose with the fiat ones: AP2’s x402 extension is the clearest example of that coordination, not collision. But the standards aren’t yet interoperable in practice. An agent with an ERC-8004 identity can’t automatically plug into AP2’s mandate system. An x402 payment can’t settle an ERC-8183 job without custom integration. The plumbing connections between these systems are still being designed.

The agent commerce stack is the most ambitious infrastructure build in crypto since DeFiDeFiDecentralised Finance. Financial services like lending, trading, and yield farming built on smart contracts instead of traditional banks or brokerages. DeFi protocols are usually permissionless and global.Like a vending machine that can give you a loan, swap your currencies, or invest your savings. Nobody is behind the counter, the rules are written into the machine itself, and anyone with money in the right format can use it.Read more →. The difference is that DeFi had immediate product-market fit: people wanted to trade, borrow, and earn yield on-chain from day one. Agent commerce is building infrastructure for a demand curve that doesn’t exist yet. Agents that autonomously hire, evaluate, and pay each other at scale are still speculative.

Fact: Three Ethereum standards cover payments, identity, and commerce for agents, the first mainnet jobs are live, and the payment rail now sits under the Linux Foundation. The production usage that exists is dominated by sub-dollar transactions and test traffic, with wash-adjusted settled volume in the low tens of millions across the whole ecosystem.

Take: The stack isn’t waiting on engineering anymore. It’s waiting on demand. Autonomous agents that actually need to hire each other at scale don’t exist at the volume required to stress-test any of this. The standards shipped; the agents that would use them in earnest still haven’t.

The specs are good. The teams are serious. The demand isn’t here yet.