Google's A2A specification v1.0 still shows a sample TextPart whose text is "Hello, world!" The transport is JSON-RPC 2.0 over HTTPS. The sample is a sentence a person could have typed.
MCP and A2A already know how to send JSON. I covered those protocols in Towards the Agentic Web. The official sample is still English.
The sample is still Hello, world
A2A's main method is message/send. A message has a role and a list of parts. A part can be text, a file, or data. Google's ADK walkthrough names the two that matter: TextPart for English, DataPart for JSON.
Older drafts tagged a part with "kind": "text". Version 1.0 switched to a member name: { "text": "Hello, world!" } or { "data": { ... }, "mediaType": "application/json" }. The greeting survived. There is still no shared list of nouns.
The Agent Card lives at /.well-known/agent-card.json. Name, skills, endpoint, auth. Discovery is a JSON file. The A2A GitHub samples still include a greeting.
MCP, released by Anthropic in November 2024, is the other protocol. The live TypeScript schema, as of 28 July 2026, is in the MCP specification repo. A tool is a function with a JSON Schema for its arguments. tools/list returns objects. tools/call returns objects. English can sit in a description so a model can pick the tool. The call itself is not a paragraph.
In December 2025 Anthropic put MCP under the Agentic AI Foundation at the Linux Foundation, with Block and OpenAI. MCP is agent-to-tool. A2A is agent-to-agent. The A2A site repeats that split. Neither protocol bans English. Both make it optional.
IBM's Agent Communication Protocol used REST and OpenAPI instead of JSON-RPC. On 27 August 2025 IBM archived the ACP repo and folded it into A2A. A2A v1.0 still lists HTTP/REST and gRPC as other ways to send the same DataPart. Merging ACP did not decide what goes inside the part.
FIPA already tried this
In the 1990s, KQML and then FIPA ACL tried to make software agents talk in a shared language. A FIPA message named the act (inform, request, refuse) and put the content in a formal language. Two agents had to agree, in advance, on the nouns.
The open web never agreed on the nouns. XML and then JSON became the payload. REST became the transport. Language models then made English cheap: say what you want and hope the other model parses it. That is why the A2A sample looks like a chat box.
What survived from that era is smaller. Schema.org for search. OpenAPI for humans who write clients. Agent Cards are OpenAPI for agents. DataParts are JSON Schema with extra wrapping. Linux Foundation AAIF can host MCP and A2A. It cannot make price mean the same thing at Amazon and a Delhi pharmacy.
The FIX Protocol has moved orders since the 1990s. A New Order Single is tags: 35=D, 55=AAPL, 54=1, 38=100. SWIFT went from MT messages to ISO 20022 XML. Nobody at a matching engine wants "please buy some Apple, not too expensive."
Google's Agent Payments Protocol (AP2) follows that rule. A Mandate is a signed object that says what the user allowed. x402 reuses HTTP 402: the agent asks, the server answers 402, the agent pays, the agent retries. On 14 July 2026 the Linux Foundation announced the x402 Foundation, with Coinbase contributing the protocol and 40 named members including Visa, Mastercard, Stripe, AWS, and Google. The project's homepage that day showed 75.41 million transactions and $24.24 million volume in the last 30 days. That is the project's meter, not an audit.
A buying agent that reads price_usd, in_stock, lead_time_days, return_window is an agent you can test. A buying agent that reads brand copy is an agent you can fool. I already put the market in agentic commerce.
Function calls already use fields
OpenAI shipped function calling in 2023, then structured outputs in 2024. Anthropic's tool use and Google's Gemini function calling do the same job. None of those objects contain a greeting. The greeting appears when a developer adds a system prompt.
MCP lets any client find a function. A2A lets any agent ask another agent to run one. That is RPC with extra names.
LangGraph, AutoGen, CrewAI, and Google's ADK still pass a lot of strings between nodes, because the node is a language model. Two models on a loop write each other essays. A schema in the middle makes them fill fields.
On 14 August 2026, OpenAI's API pricing still billed input and output by the million tokens. A long English plan can be thousands of tokens. A DataPart can be eighty.
A buyer agent wants a flight. In English: "Find a refundable economy seat SFO to DEL leaving 12 September, back 28 September, under 1200 dollars, aisle if cheap." "Aisle if cheap" has no test. In JSON: {"origin":"SFO","dest":"DEL","depart":"2026-09-12","return":"2026-09-28","cabin":"Y","max_usd":1200,"refundable":true,"seat":"aisle","seat_flex_usd":40}. IATA NDC already pushed airlines toward XML offers for this reason.
Healthcare has HL7 FHIR. An agent that summarises a chart in English, then a second agent that acts on the summary, is two guesses stacked. A FHIR Bundle is the DataPart. English is the note for the human.
An Agent Card can list a URL. It cannot prove who paid for the agent. AP2 mandates, W3C verifiable credentials, and OpenID name a principal. A paragraph that says "I am Alice's agent" is phishing.
Cisco's AGNTCY, donated to the Linux Foundation in July 2025, is a directory, a schema, and a transport. It is not a message format. MCP, A2A, ACP, AP2, and AGNTCY do different jobs.
What a person can no longer read
A regulator can read a TextPart. A DataPart is readable if you know the schema. An embedding is not. If two agents start passing vectors, the log is a tensor. You can store it. You cannot read it.
Your Order and my Order will drift. Agent Cards can list skills. They cannot force two vendors to mean the same thing by refund. ISO 20022 works because banks were forced. Agent markets are not.
When a travel agent hires a fare agent, and the fare agent hires a payments agent, your screen is a summary of a job you did not see. That is acceptable for a flight. It is a problem for a medical prior auth or a parole report.
Courts still want sentences. Soelberg, the Times log order, every discovery fight. A vendor that stores only vectors will meet a judge who asks for the emails. The vendor will then write a story after the fact.
The EU AI Act Article 50 is a labelling rule. NIST's AI Risk Management Framework asks for an audit trail. A TextPart looks like "what the agent said." A DataPart is the object. A vector is neither unless you attach a decoder. If procurement asks for the English form, engineers will still send JSON. You will pay for both.
JSON-RPC is a community spec, not an IETF RFC the way HTTP is. gRPC is a CNCF project. A2A is a Linux Foundation project run by cloud vendors. That is enough to ship. A country that wants an auditable log will have to write it into law, the way PSD2 wrote bank APIs, or accept whatever the vendors keep.
Production agents still send TextParts. The test is the first high-value job that finishes with a Card, a DataPart, a Mandate, a receipt, and no hello. People will still type English at the keyboard.