Your AI agent is about to start buying things without you

4 min read

The infrastructure for AI agents to buy and sell from each other is being built right now. Not in a lab. In production.

Four protocols have emerged that cover every layer of agent communication. The IMF published a note on how agentic AI will reshape payments. Google has 50-plus partners on its agent-to-agent protocol. IBM contributed a commerce protocol to the Linux Foundation.

This is not a product announcement. It is infrastructure. And it moves faster than most people realize.

The protocol stack

In 2024, every agent framework had its own tool-calling convention, its own coordination mechanism, and its own approach to transactions. It was chaos. You picked a framework and you were locked into its way of doing everything.

By early 2026, four protocols have carved up the stack:

  • MCP (Model Context Protocol) — Anthropic’s standard for connecting agents to tools, APIs, and data. 97 million downloads. Adopted by OpenAI, Google, and Microsoft. It won the agent-to-tool layer.
  • A2A (Agent-to-Agent Protocol) — Google’s answer for agent coordination across vendor boundaries. Agents discover each other’s capabilities and delegate tasks. 50-plus launch partners. Built-in security via Agent Cards.
  • ACP (Agent Commerce Protocol) — IBM’s open standard for agent-to-agent commercial transactions. Contributed to the Linux Foundation. Handles payment, fulfillment, and transaction semantics.
  • UCP (Universal Commerce Protocol) — Google’s commerce-specific protocol for agents operating inside Google’s ecosystem.

These do not compete. They compose. A complete agent stack in 2026 uses all four: MCP for tool access, A2A for coordination, ACP or UCP for commerce.

Why this matters

The shift is subtle and consequential. Up until now, agents have been tools that act on behalf of humans. A research agent reads papers. A coding agent writes code. A booking agent finds flights. The human is still in the loop for every transaction that involves money.

Agent-to-agent commerce changes the loop. An agent can now research a product, compare prices, negotiate terms, and complete a purchase — all without a human approving each step.

The IMF’s note on agentic AI and payments is not theoretical. It covers cross-border payment routing, compliance, and settlement for transactions where neither party is human. That is the kind of thing central banks only write about when it is already happening.

The invisible shift

This is the part that gets missed because it is boring. Nobody is demoing an agent buying cloud credits from another agent at a keynote. The real work is in the protocol specs, the Agent Card formats, the payment semantics, the compliance layers.

But boring infrastructure is what enables everything else. HTTP was boring. JSON was boring. Nobody gave a TED talk about them. They became the substrate that built the modern internet.

Agent protocols are becoming the substrate for the next layer. An infrastructure where software agents are the primary actors in commercial transactions.

The risk nobody is talking about

The permission problem is still unsolved. If an agent has a payment method and the authority to spend, what stops it from spending too much? From buying from the wrong vendor? From being manipulated by a malicious agent into a bad deal?

A2A has Agent Cards — structured capability declarations that let agents verify what other agents can do before delegating to them. That is a start. But capability verification is not the same as spend authorization.

The companies building these protocols are solving the coordination problem. The authorization problem — who gets to spend what, and how do you enforce it — is still largely an afterthought.

What to watch

The protocol convergence is the signal. When four competing standards stop competing and start composing, the market has decided. The agent communication layer is solidifying.

Three things to track:

  • Adoption beyond the Big Four. MCP has 97 million downloads, but most of that is Anthropic-driven. Real adoption is when a mid-market SaaS company exposes its API as an MCP server because that is what its customers expect.
  • Compliance frameworks. The IMF note is the first shot. Financial regulators will follow. The companies that build compliance into their agent stacks from day one will win the enterprise deals.
  • Agent marketplaces. Once agents can transact with each other, someone will build the exchange. Not a human-facing store. An API-first marketplace where agents discover, negotiate, and transact. That platform will capture more value than anyone expects.

The bottom line

The conversation about AI agents has been stuck on capability. Can it code? Can it browse? Can it hold state? Those questions are answered. The interesting question now is what happens when agents start acting as economic actors — buying services, negotiating contracts, settling payments.

The protocols are here. The IMF is writing about it. The infrastructure is being laid down.

Your agent is not just going to do your work. It is going to spend your budget.

The only question is whether you set the limits before it starts buying.