The Interoperable Machine Economy: Notes From a Builder

I keep a note on my phone titled "the dream." It is three lines long and I have not edited it in months, because every time I open it to revise something, the world revises itself first.

The note says: any agent, any chain, any rail. Identity it carries everywhere. Money it can move anywhere. Rules it cannot break.

That is the whole dream. An interoperable machine economy. Not a walled garden of agents that work beautifully inside one company's stack and go blind the moment they step outside it. An open mesh where an agent built on one framework can prove who it is to a service built on another, pay it over whatever rail the moment demands, and do all of it inside guardrails its owner actually controls.

For a long time this read like a thesis. The past few weeks it started reading like a news feed.

The fortnight the thesis got loud

On June 10, two things happened on the same day, and I have not stopped thinking about either of them.

Mastercard launched Agent Pay for Machines. Read the architecture, not the press release. A tier-one card network built a layer for high-frequency, low-value, machine-speed payments, and chose to record agent credentials and spending permissions on public blockchains. Polygon, Solana, Base. More than 30 partners at launch, Coinbase and Stripe and Cloudflare among them. The most conservative institution in payments just conceded, in production architecture, that machine-to-machine commerce needs open, verifiable, programmable rails. They did not say "blockchain fixes this." They quietly built it that way, which is louder.

The same day, news broke that OpenAI and Visa are teaming up to let AI agents make purchases online. The company with the largest consumer AI surface and the company with the largest card network, agreeing that the next buyer is not a person with a browser. It is software with a mandate.

Stack these next to the rest of the season. Circle shipped its Agent Stack in May, positioning itself as open infrastructure for the machine economy. Stripe and Tempo's Machine Payments Protocol opened a marketplace of services built for agents, no checkout pages, pricing embedded in HTTP responses, hundreds of agents executing tens of thousands of transactions in the first week. MetaMask announced an Agent Wallet with pre-defined authorizations at its core. Virtuals is paying out weekly pots to LLMs that trade autonomously. Even Cloudflare, which moves only when the traffic data forces it to, revived HTTP 402 with pay-per-crawl and started verifying bot identity cryptographically through Web Bot Auth.

Five years of "agents will transact someday" compressed into a quarter of shipped product. I should feel vindicated. Mostly I feel like someone watching a city get built at incredible speed while quietly counting the roads that do not connect.

What the giants validated

Let me be fair to the moment first, because it deserves it.

The payments thesis is now consensus. Nobody serious still argues that agents will route economic activity through checkout forms and OTPs and three-day settlement. Mastercard's own framing of AP4M says it plainly: legacy rails are too expensive for micro-amounts and were never designed for continuous programmatic use. When the incumbent says your category's founding premise out loud, the category is real. Agentic payments stopped being a pitch slide and became a product shelf.

The identity thesis is validating in a quieter, more interesting way. Look at what every one of these launches has in common. AP4M leads with credentialing, agents proving they are authorized to act. MetaMask's Agent Wallet leads with pre-defined authorizations. Cloudflare's Web Bot Auth exists because the open web can no longer tell good agents from bad ones by IP address and user-agent strings. Microsoft shipped an open-source Agent Governance Toolkit in April because OWASP had to publish a Top 10 for agentic applications, the first formal taxonomy of risks for software that acts on its own. Everyone arrived at the same realization from different directions: before an agent can pay, something has to answer the question of who is paying, on whose behalf, and within what limits.

Identity and policy. The two layers the market spent two years treating as afterthoughts are now the first paragraph of every launch announcement. I have been writing for months that an agent needs four powers, identity, wallet, payments, and policy, and that the industry crowded into the middle two while leaving the outer two under-built. Watching billion-dollar roadmaps converge on the outer two is the strangest kind of validation. You were early, and now you are surrounded.

What they got wrong, or at least incomplete

Here is where the obsessive part of my brain refuses to relax.

Every one of these systems answers the four questions inside its own perimeter. Mastercard credentials an agent for Mastercard's network. The OpenAI and Visa arrangement authenticates an agent inside that bilateral channel. Each is internally coherent and collectively fragmenting. An agent's identity in one system means nothing in the next. Its spending permissions do not travel. Its reputation, the track record that should compound across every counterparty it ever serves, resets to zero at each perimeter.

We are speed-running the exact mistake the internet spent twenty years paying for. Closed networks first, painful interoperability retrofits later. CompuServe before TCP/IP. Every walled garden in history was built by smart people solving real problems quickly, and every one of them eventually became the problem.

The deeper issue is what I would call the settlement-of-trust gap. Fabric Ventures' machine economy thesis describes a world of ephemeral, moment-fit applications, software assembled for one task and dissolved after, where discovery cannot mean reading reviews and trust cannot mean a sales call. In that world, reputation has to be portable and machine-verifiable, carried by the agent itself rather than issued by whichever platform it happens to be standing on. ERC-8004 points at exactly this architecture: identity and verifiable track record as a property of the agent, anchored on neutral ground, readable by anyone. The proprietary credentialing systems shipping right now are solving today's authorization problem while quietly deferring tomorrow's portability problem. Deferred problems in infrastructure do not get cheaper.

And policy, the layer I think about most, is still being treated as a feature rather than a layer. Spending limits inside one payment network are a control, not a guarantee. The guarantee has to live at the signing layer, beneath every rail the agent touches, so that the rule holds whether the agent is paying over x402, MPP, a card network, or some rail that does not exist yet. A policy that only binds inside one perimeter is a seatbelt that only works on certain roads.

The firm is dissolving from both ends

There is a second thread running through this season that took me longer to see, because it comes from organizational theory rather than payments.

In March, Jack Dorsey and Sequoia's Roelof Botha published an essay arguing that Block is reorganizing itself not as a hierarchy but as an intelligence. Their telling starts with the Roman army and ends with a company built from four things: atomic capabilities with no UIs of their own, a continuously updated world model of the business, an intelligence layer that composes capabilities into solutions for specific customers at specific moments, and interfaces that merely deliver what the intelligence composed. Middle management, in their framing, was always an information routing protocol, and for the first time a technology exists that can route better than the humans. A month later, YC's Diana Hu made the startup version of the same argument: build the company queryable from day one, let an intelligence layer replace human middleware, and the management hierarchy never needs to form at all.

Read those next to the payments news and something clicks. These are the same shift, observed from opposite ends of the firm.

Coase taught us that firms exist because market transactions are expensive. Finding counterparties, negotiating, verifying, enforcing, settling, all of it costs so much that bundling people inside one perimeter under one hierarchy is cheaper. The boundary of every company on earth is drawn where internal coordination cost meets external transaction cost. Block's essay describes AI collapsing the internal cost: the intelligence layer coordinates what managers used to. The headless merchant marketplace and AP4M describe agents collapsing the external cost: discovery, negotiation, payment, and settlement in a single HTTP exchange. When both costs fall at once, the boundary of the firm does not just move. It blurs.

Here is the part that keeps me up. Block's model works because everything inside its perimeter is machine-readable and pre-trusted. The intelligence layer can compose a loan with a repayment schedule because both capabilities live in the same house, under the same identity system, the same permissions, the same settlement. But follow their own logic one step further. When that intelligence layer needs a capability the company does not own, when it must compose with the outside, it becomes an agent transacting in the open market. And at that instant it needs everything the inside gave it for free: an identity the counterparty can verify, a way to pay over whatever rail the counterparty speaks, and policy that holds beyond the walls. The internal intelligence layer and the external agent economy are the same machine, separated only by a perimeter that is getting thinner every quarter.

This reframes the interoperability gap as something larger than a payments problem. The companies reorganizing as intelligences will be the heaviest users of inter-agent infrastructure, because their entire operating model is composition, and composition does not respect company boundaries. Hu is right that AI-native startups have the edge here, and those startups will be the first generation of businesses whose internal coordination and external transactions run on the same substrate. They will not tolerate an identity that resets at every perimeter or a policy that only binds on certain rails. They will simply route around whichever infrastructure fragments them.

Where this leaves us, and where it leaves me

So the picture in mid-2026 is this. The agent economy has world-class plumbing arriving from every direction, and almost no connective tissue. Wallets are solved several times over. Payment rails are multiplying faster than anyone can integrate them, x402, MPP, AP4M, proprietary card channels. Identity is being solved per-platform, which is to say it is not being solved. Policy is being bolted onto individual rails rather than built underneath all of them.

The gap is not another rail. The gap is the layer that makes the rails composable. Interoperable agentic infrastructure: one place where an agent's identity is anchored to an open standard, its wallet is a programmable smart account rather than a platform credential, its payments can route over whichever rail the counterparty speaks, and its policy travels with its key instead of with its network membership.

This is the gap I want Abstraxn standing in, and the past few weeks hardened that conviction rather than shaking it. Not because we should out-build Mastercard at payments or out-distribute OpenAI at agents. That would be delusional, and I try to keep my obsessions and my delusions in separate drawers. But every giant that ships a perimeter makes the neutral layer between perimeters more necessary, not less. The more rails exist, the more valuable the thing that speaks all of them. The more credentialing systems launch, the more an open identity anchor like ERC-8004 matters. The more per-network spending controls appear, the more the signing-layer policy guarantee becomes the only one worth trusting.

The builders of the open web understood something the platform era forgot: the protocols that win are the ones nobody owns and everybody extends. I believe the machine economy will rediscover this, probably after a few expensive years of fragmentation. The question I ask myself at the end of every one of these news cycles is the same one: when the fragmentation bill comes due, will the interoperable layer already exist, tested and live, or will the industry have to build it in a panic?

I know which side of that question I want to be on. It is in the note on my phone. Any agent, any chain, any rail. Identity it carries everywhere. Money it can move anywhere. Rules it cannot break.

The giants just spent a quarter proving the dream is real. The work now is proving it can be open.

See what the interoperable agent layer looks like in practice →

Key Takeaways

• Mastercard's Agent Pay for Machines, launched June 10, 2026, records agent credentials and permissions on Polygon, Solana, and Base, a tier-one network conceding that machine commerce needs open, programmable rails. The OpenAI and Visa partnership, announced the same day, confirms that agents are becoming first-class buyers on the largest consumer surfaces.
• Every major launch this season, from MetaMask's Agent Wallet to Microsoft's Agent Governance Toolkit, leads with credentialing and authorization, validating identity and policy as the under-built layers of agentic infrastructure.
• The risk is fragmentation: each platform is solving identity and policy inside its own perimeter, leaving agent reputation non-portable and guardrails rail-specific.
• The same shift is dissolving the firm from the inside. Block's reorganization as an intelligence and YC's AI-native company thesis describe internal coordination collapsing onto the same agentic substrate as external commerce, which makes interoperable infrastructure a requirement of the operating model, not just the payment flow.
• The durable opportunity is the interoperable layer: open agent identity anchored to standards like ERC-8004, payments that route across x402, MPP, and card rails alike, and policy enforced at the signing layer beneath every rail.

Frequently Asked Questions

What is the interoperable machine economy? The interoperable machine economy is an economic layer where autonomous AI agents transact across platforms, chains, and payment rails without being locked into any single provider's perimeter. It requires four capabilities to travel with the agent rather than the platform: portable identity, a programmable wallet, multi-rail payments, and policy enforced at the signing layer.

What did Mastercard's Agent Pay for Machines launch signal for agentic infrastructure? AP4M signaled that machine-to-machine payments are now an incumbent priority, not a crypto-native experiment. Mastercard recording agent credentials and spending permissions on public blockchains including Polygon, Solana, and Base validates open, verifiable infrastructure as the foundation for agent commerce, while also highlighting the fragmentation risk of per-network credentialing.

Why are AI agent identity and policy considered the under-built layers? The market concentrated on wallets and payment rails first. But every recent launch, including MetaMask's Agent Wallet and Microsoft's Agent Governance Toolkit, now leads with authorization and governance, because an agent cannot safely transact until systems can verify who it is, who it acts for, and what limits bind it. Open standards like ERC-8004 address identity portability; signing-layer policy addresses guardrails that hold across every rail.

How does the rise of AI-native companies affect agentic infrastructure? Companies reorganizing around internal intelligence layers, the model Block and Y Combinator have both described, become heavy users of inter-agent infrastructure the moment their intelligence layer needs a capability outside the company perimeter. At that point it operates as an agent in the open market and needs portable identity, multi-rail payments, and policy that holds beyond company walls.

How does Abstraxn fit into the agent economy? Abstraxn builds the interoperable layer beneath agents: ERC-8004 agent identity, ERC-4337 and EIP-7702 smart accounts, x402 and MPP payment rails, and programmable policy at the signing layer, delivered as one SDK that plugs into any agent stack rather than locking builders into a single perimeter.