On April 20, Cloudflare announced that AI agents can now create accounts, start subscriptions, and deploy code without a human in the loop. Stripe, powering the payment layer underneath, sets a default spending limit of $100 per month per provider.
An agent has a credit limit. Let that register.
This is one of at least three independent efforts building the same kind of thing. In February, Coinbase released Agentic Wallets, infrastructure for AI agents to spend, earn, and trade autonomously within programmable guardrails. The human sets policies upfront (session caps, transaction limits, allowlisted contracts), and the agent operates inside that envelope. Meanwhile, competing commerce protocols have emerged from different consortia, each defining how agents discover products, negotiate terms, and complete purchases. Consider who's endorsing this work: Visa, Mastercard, Target, Walmart, and over twenty other partners have signed onto a single protocol governing how agents buy things. The payments infrastructure of the consumer economy is being retooled, in public, for participants that don't eat.
None of these companies coordinated with each other. They arrived at the same plumbing from different starting points, solving different immediate problems. When that happens, the convergence usually matters more than any individual piece.
And what's converging has started to look like something beyond agents performing human tasks faster. On-chain identity standards like ERC-8004, live on Ethereum's mainnet since January, give agents portable, verifiable identities. Commerce protocols give them a language for transactions. Wallet infrastructure gives them the ability to hold and spend money. Each piece is pragmatic on its own. Together they compose something that starts to resemble economic standing, though we don't have good language for what economic standing means when the entity holding it has no preferences and no capacity to want anything. We're building the infrastructure before we've figured out what to call the thing it enables.
We have a partial precedent. In 1800, the United States had roughly 300 corporations. By 1900, approximately 500,000. That explosion didn't happen because someone declared corporations to be people. It happened because infrastructure accumulated: the right to hold property, then to enter contracts, then to sue and be sued. General incorporation laws in the 1840s removed the need for special legislative grants, letting anyone who met basic standards form a corporation. Each step was a practical fix for a specific friction. The philosophical implications arrived well after the plumbing was in place.
The current moment rhymes: infrastructure is arriving before the framework for thinking about it. But corporations gained their standing through legislatures, courts, and centuries of deliberation. Agents are gaining theirs through product launches and protocol specifications, on timelines measured in months.
Web identity was never resolved by design. It defaulted to a handful of platforms that moved faster than governance, and then everyone lived inside that resolution as though it had been chosen. The people building agent commerce infrastructure are trying to ship features. But the aggregate effect of those features, piece by pragmatic piece, is assembling a new class of economic actor in the ordinary course of solving engineering problems.
Whether there's still time to be deliberate about what's being assembled remains genuinely unclear. The history of defaults suggests the window is smaller than it looks.
-
NIST enters the frame: The U.S. government launched its AI Agent Standards Initiative in February 2026, the first federal program dedicated to interoperability and security standards for agentic systems, with a public comment period on agent identity and authorization that closed in April.
-
The naming collision problem: Two entirely unrelated protocols both called "ACP" are circulating — IBM's agent communication standard (now merged into Google's A2A) and OpenAI/Stripe's agentic commerce protocol — and sources routinely conflate them, which matters as enterprises try to build coherent agent stacks.
-
Amazon draws a line: A federal court blocked Perplexity's Comet agent from accessing Amazon accounts, ruling that user permission and platform authorization are legally distinct — a precedent that could shape which agents get to participate in commerce and which get locked out.
-
Only 14% reach production: Multiple surveys converge on the same finding: while roughly 78% of enterprises are piloting AI agents, no more than 14% have reached production scale, suggesting the economic infrastructure being built for agents is running well ahead of the organizations that would govern their use.