Your AI agents can now buy things — and the rails to do it safely just shipped.

For two years the knock on AI agents was that they could talk but couldn’t do. That objection just collapsed. In a single stretch of 2026, agents stopped being advisors and became actors that move money — and, crucially, the rails arrived with the trust controls already in them.

• Visa completed hundreds of secure, agent-initiated transactions with partners across its network — real purchases, completed by agents, at Visa scale.
• On June 10, Mastercard launched Agent Pay for Machines — programmatic, high-velocity, low-value payments between agents — backed by 30+ partners including Stripe, Adyen, Coinbase, Cloudflare and Checkout.com. Its chief product officer called it a potential “superbloom of AI business models.”
• The Coinbase-led x402 protocol has already processed ~165 million agent transactions.
• Juniper Research sizes agentic commerce at $8 billion in 2026, scaling to $1.5 trillion by 2030 — with McKinsey’s range running to $3–5 trillion globally.

This is the unlock. An agent that can read your knowledge base is useful. An agent that can complete a transaction — reorder inventory, settle a microtransaction, pay an API per call, close a purchase — is a new category of software. And the most important detail isn’t the dollar figures. It’s how these rails were designed.

The rails came with the controls built in

Look at what Mastercard actually shipped. Agent Pay for Machines isn’t just a faster pipe — it’s credentialing and permissioning: an agent has to prove it’s authorized to act, and users set programmatic spending rules and limits that are enforced automatically, transaction by transaction. Visa’s pitch leads with “secure” for the same reason. The payment networks understood something important: agents that spend money are only valuable if you can bound exactly what they’re allowed to spend it on.

That’s not a footnote — it’s the whole design. The opportunity and the guardrail shipped as one product, because one is worthless without the other. Juniper’s own report names trust as the number-one factor that decides who captures this market. The teams that win the $1.5 trillion won’t be the ones who move fastest or slowest on agents — they’ll be the ones whose agents are scoped, permissioned, and auditable enough to turn loose with a credit line.

Why this is good news, not scary news

Here’s the optimistic read, and it’s the correct one: the hard part is now standardized. You no longer have to invent agent identity, spending limits, or transaction audit from scratch — Visa, Mastercard, and the protocol layer just built that floor for everyone. What’s left is the part that was always the real work and the real margin: defining what your agent should do, scoping what it can touch, and measuring whether it’s doing it well.

That maps exactly onto the discipline we already build every system on — scoped, least-privilege credentials; hard spending and rate limits; a kill switch that cuts access in a single call; and a telemetry record of every action. Until this month that was the operational layer that kept a retrieval agent safe. Now it’s the same layer that lets a transacting agent be turned on with confidence. The pattern didn’t change. The stakes — and the upside — just got bigger.

What to do this week

1. Pick one bounded, repetitive purchase or payment your business makes constantly — supplier reorders, per-call API spend, subscription top-ups. That’s where the first safe, high-ROI transacting agent lives.
2. Write down the spending envelope before you build: the per-transaction cap, the daily ceiling, the exact merchants or APIs it may pay, and the human-approval threshold. The new rails enforce these for you — but only if you’ve decided them.
3. Instrument the audit trail from day one. Every agent payment should be logged, attributable, and reversible. With the rails handling identity and limits, this is now a configuration step, not a research project.

The bigger picture

The shift from agents that advise to agents that act on money is the biggest expansion of what this software can do since the models themselves got good. And it arrived in the healthiest possible way: with the payment networks treating scoped permissions and auditability as the product, not the compliance afterthought.

That’s the same bet we’ve made from day one — that the model is the commodity and the operational layer is the moat. This month, the largest payment networks on earth agreed with us, in production, at scale. The agents that capture the $1.5 trillion won’t be the cleverest. They’ll be the best-governed — scoped tightly enough to trust with a wallet. That’s a problem we know how to solve, and now there’s a very large reason to solve it.