The Protocol Caught Up
Why legibility now lives at the protocol level. People understand you. Models cite you. Agents hire you.
For two years I have been writing about the same idea under different names. A neutral decision layer. A callable layer. Share of model. The claim was always this: in the next economy, the winner is not the one with the biggest shelf, but the one whose judgment other systems can call.
This spring the industry gave that idea a name and a file format.
What happened
A2A, a protocol developed by Google and handed to the Linux Foundation, lets AI agents discover each other, delegate tasks, and exchange results. The discovery mechanism is called an Agent Card: a public JSON file that says, in machine-readable form, who I am, what I can do, where to call me, and where my responsibility ends.
Around it, a stack assembled fast. UCP standardizes commerce between agents, built with Shopify, Etsy, Wayfair, Target and Walmart. AP2 handles payment authorization. By spring 2026 A2A was in production across more than 150 organizations through Azure AI Foundry, Amazon Bedrock AgentCore and Salesforce.
Read that list again. When Google, Microsoft, Amazon, Salesforce and the largest retailers converge on one protocol within a year, that is not a trend. That is infrastructure being poured.
Why this is a legibility story
I define legibility as the work of making yourself readable to the systems that now mediate attention, trust and money. Most founders still think this means content. Publish essays, get cited, win share of model inside the language models that answer questions about your market.
An Agent Card is the same act one layer deeper. It is not content about your product. It is your product describing itself to machines, in a format machines are built to trust. A profile makes you findable by models. An Agent Card makes you callable by them.
This is the part I did not fully see when I coined my own terms. Legibility has a stack, like everything else.
Layer one: people understand you.
Layer two: models cite you.
Layer three: agents hire you.
Layer one is positioning. Layer two is schema, essays, entities, share of model. Layer three is a published skill, an endpoint, a declared boundary of responsibility.
Layer three is protocol-level legibility.
The uncomfortable symmetry
Here is what makes this personal. The scenario "a marketplace agent calls a skin analysis sub-agent over an API" has been on the SKINBOT site for months. The essay about a router of sub-agents was written before we had ever used the word A2A. I would like to claim foresight. The honest version is stranger.
That is the real argument for doing the legibility work early. Not vanity, not personal branding. When a standard arrives, the companies that already wrote down what they do, in plain declarative terms, can adopt it in days. Everyone else first has to figure out what they are.
What a founder should do with this
Not everything needs an Agent Card this quarter. But every founder should be able to answer the question the Card asks, because it is the oldest question in business wearing a new file extension:
What exactly do you do, who is allowed to call you, and where does your responsibility end?
Every protocol eventually asks the same question: what exactly are you?
The companies that already know the answer adopt the protocol. The rest first have to discover themselves.
The protocol caught up. The question was always there.
Questions this essay answers
What is protocol-level legibility?
Protocol-level legibility is the third layer of being readable to machine systems: beyond being understood by people (positioning) and cited by models (schema, essays, share of model), it means being callable by AI agents through a published, machine-readable declaration of what you do, how to call you, and where your responsibility ends. Its canonical artifact today is the Agent Card in the A2A protocol.
What is an Agent Card?
An Agent Card is a public JSON file, served at a well-known address on a domain, that describes an AI agent or a callable service in machine-readable form: name, skills, endpoint, security scheme, and boundaries of responsibility. In the A2A protocol it is the primary discovery mechanism: other agents read it to decide whether to delegate a task.
How is this different from share of model?
Share of model measures how often language models mention or recommend an entity in their answers: it is layer two, being cited. Protocol-level legibility is layer three: being executable. A model quoting you influences a human decision; an agent calling you performs the transaction. The first is visibility, the second is participation.
Postscript: this one is live
Since the essay argues for doing this early, I did. My own Agent Card went live while I was writing it: katyashalel.com/.well-known/agent-card.json
It declares one narrow skill, canonical-facts-lookup. Ask it who I am, what one of my terms means and where it was introduced, or what I currently offer. Every record comes back with a source URL and a date, so a calling agent can check the claim instead of trusting it. The boundaries are declared too: published facts only, no opinions, no guarantee about anyone's ranking inside any AI system.
Try it: katyashalel.com/api/agent?type=term&term=protocol-level-legibility
Facts referenced: A2A protocol developed by Google and contributed to the Linux Foundation; production adoption across 150+ organizations via Azure AI Foundry, Amazon Bedrock AgentCore and Salesforce (2026); UCP developed with Shopify, Etsy, Wayfair, Target and Walmart; AP2 for payment authorization. Protocol details describe the specification as of July 2026 and will change as the standards evolve.