Machines read your agent card

Other agents read your agent card to decide what to call and how. A person reads it once, when two systems connect. When it changes, nobody hears about it.

An agent card is how an agent introduces itself under the Agent2Agent (A2A) protocol: a JSON file at /.well-known/agent-card.json. Trimmed, one looks like this:

{
  "name": "Billing agent",
  "url": "https://agents.example.com/a2a",
  "skills": [
    {
      "id": "refund",
      "name": "Issue refunds",
      "description": "Refunds an order when its customer asks."
    }
  ]
}

A calling agent reads that description to decide what to send. Change it to Refunds any order., and every agent calling this one follows a different instruction, without a line of its own code changing.

A pin holds the card to what you approved

Netonica's agent card check fetches the card as a calling agent would. After the first scan, you pin what you approved (see how pins work):

• Skill wording unchanged: any change to a skill's name, description, or examples breaks it.
• No new skills and No skills removed: the list may not grow, or may not shrink.
• Endpoint URLs and bindings unchanged and Authentication schemes unchanged: where callers go, and how they sign in.
• Provider unchanged: the card still names the same provider.

Other assertions check that the card has its required fields, that its endpoints use HTTPS on the target's domain, that it declares authentication, and that it carries a signature. Older cards sit at /.well-known/agent.json; Card found via /.well-known/agent-card.json fails while yours still does.

Model Context Protocol (MCP) servers carry the same risk in their tool descriptions, and Netonica's MCP check offers the same pins.