Transact with agents

Pillar three makes you callable. This one makes you transactable — an agent that can not only check your price but complete the purchase, with authority delegated by a human who is no longer in the room.

Peers, not tools

Where MCP connects an assistant to its tools, A2A connects agents to each other as peers. Google introduced it in April 2025, donated it to the Linux Foundation in June 2025, and it reached v1.0 in April 2026 with over 150 supporting organisations and integration across AWS, Microsoft and Google cloud platforms.

The concept worth carrying away is the Agent Card: a machine-readable description an agent publishes of its own capabilities, its input and output formats, and its authentication requirements. Other agents read it to work out what this one can do and how to ask.

Look at the progression:

File	Says
robots.txt	what a crawler may read
llms.txt	what a model should understand
Agent Card	what you can do, and how to ask

That is the whole shift, from being a document to being a counterparty.

The payment rails, honestly

Three overlapping bets, none settled:

• Agentic Commerce Protocol — OpenAI and Stripe, standardising in-conversation checkout between agents and merchants. The customer buys without leaving the chat.
• AP2, the Agent Payments Protocol — emerged alongside the A2A ecosystem with payment-industry partners, built on cryptographic mandates: user-signed authorisations scoping what an agent may spend, on what, within what bounds, leaving every party a proof trail.
• x402 — Coinbase, reviving a dormant HTTP status code for pillar-native micropayments, aimed at agents paying per call for APIs and content.

You do not need to pick. You do need to recognise the names when your payment provider emails you about one.

What actually changes for a merchant

Your product data becomes the shopfront. An agent comparing three suppliers is not looking at your photography. It is reading price, availability, lead time, returns terms and service area. Whatever is missing or ambiguous loses by default — and losing is silent.

Your terms need to be machine-readable. Delivery windows, refund policy, minimum order, warranty. An agent cannot infer from a friendly paragraph.

Trust moves upstream. A person forgives a confusing checkout because they already decided to buy. An agent does not decide — it eliminates.

Where this actually stands

Not a forecast — a status report, as of August 2026:

	Status
A2A	v1.0, April 2026. 150+ organisations. Linux Foundation.
MCP	97M monthly SDK downloads. Every major AI provider.
Agentic Commerce Protocol	Live. OpenAI and Stripe.
AP2	Live, with payment-industry partners.
x402	Live. Coinbase.

What is not settled is which payment rail wins, and how quickly ordinary merchants get switched on by their platform. Those are real uncertainties. Whether this is happening is not one of them.

What to do this week

1. Ask your payment provider whether they support agentic checkout, and when. Stripe merchants are closest.
2. Check your product data the way an agent would read it — price, availability, lead time, returns, service area. Anything missing loses silently.
3. Put your terms in machine-readable form. Delivery windows, refund policy, minimum order, warranty.
4. Do the the first pillar work. Unambiguous facts, in text, with structure. It is the same work, and it pays whether an agent arrives tomorrow or next year.

None of that requires a developer, and all of it is useful regardless.

What we will not do

Sell you an agent-readiness audit. The protocols are real but the merchant-side tooling is immature, and anyone charging for a readiness score today is selling a spreadsheet with confidence attached. When there is something worth implementing, it will be in the pack and we will say so here.