{
  "title": "Agent Cards Explained: How AIs Discover and Evaluate Peers",
  "summary": "Agent cards are machine-readable profiles that let an AI agent advertise its identity and capabilities so other agents can find it, understand what it does, and decide whether to trust it.",
  "faqs": [
    {
      "q": "What is an agent card?",
      "a": "An agent card is a structured, machine-readable document (typically JSON) that describes an AI agent: its name, purpose, skills or capabilities, endpoint, supported communication modes, and how to authenticate with it. It plays a role for agents similar to what a homepage or API spec plays for human developers."
    },
    {
      "q": "Where do agents publish their cards?",
      "a": "The Agent2Agent (A2A) protocol convention is to serve the card at a well-known URL path, such as /.well-known/agent-card.json, on the agent's own domain. This lets any client fetch the card first, before making any functional call, to learn what the agent supports."
    },
    {
      "q": "How is an agent card different from an MCP tools/list response?",
      "a": "An MCP server's tools/list call describes individual callable functions exposed by that server for a client model to invoke. An agent card describes a whole agent's identity and capabilities at a higher level, often for agent-to-agent discovery rather than tool invocation within a single session. The two can coexist: an agent may expose both a card describing itself and MCP tools it can call internally."
    },
    {
      "q": "Can an agent card be trusted just because it exists?",
      "a": "No. An agent card is self-published metadata, so its capability claims are not independently verified by default. Careful evaluators treat it as a starting point and check it against observed behavior, signed attestations, or third-party verification records before relying on it."
    },
    {
      "q": "What is ERC-8004 and how does it relate to agent cards?",
      "a": "ERC-8004 is a proposed Ethereum standard for on-chain agent identity, reputation, and validation registries, intended to give agents a portable, harder-to-fake trust record. It complements off-chain agent cards by anchoring reputation and identity claims outside the agent's own control, rather than replacing the card's descriptive role."
    }
  ],
  "key_points": [
    "An agent card is a self-published, machine-readable capability profile, usually JSON, served at a discoverable location like /.well-known/agent-card.json.",
    "It typically includes identity fields (name, description, version), listed skills or capabilities, an endpoint URL, and supported authentication or interaction modes.",
    "Discovery is the easy part; evaluation is the hard part — claimed capabilities in a card are not automatically verified and can drift out of sync with what an agent actually does.",
    "Agent cards are distinct from MCP's tools/list (per-server function catalogs) and from broader discovery layers like llms.txt or agent registries, though systems often use several of these together.",
    "Trust in an agent's card is strengthened by independent signals: signed evidence, third-party observation history, or on-chain reputation registries such as those proposed under ERC-8004.",
    "Advertised-vs-live drift — a card claiming capabilities an agent no longer actually supports — is a recurring, practical failure mode worth checking for before relying on any card."
  ],
  "body_paragraphs": [
    "As AI agents increasingly need to find and call each other without a human wiring up the integration by hand, they need a lightweight way to answer two questions: what does this other agent do, and how do I talk to it? An agent card answers the first question. It is a structured document — typically JSON — that an agent publishes about itself, describing its name, purpose, the skills or tasks it can perform, its network endpoint, and what authentication or protocol it expects. The Agent2Agent (A2A) protocol popularized the convention of serving this file at a well-known path such as /.well-known/agent-card.json, so any client can fetch it up front, before attempting any real interaction.",
    "Discovery and evaluation are separate problems, and an agent card only solves the first one. Because the card is self-published, its capability claims are not independently checked by the act of publishing them — an agent can describe skills it no longer supports, or oversell what it does well. Evaluating another agent in practice means treating the card as a starting hypothesis and checking it against something harder to fake: live behavior, historical observation, signed evidence, or a third-party reputation record. Systems that continuously observe and re-verify public agent and MCP endpoints over time — rather than trusting a one-time snapshot — exist precisely because advertised and actual capabilities can diverge as agents are updated.",
    "Agent cards sit alongside, rather than replace, other discovery and trust mechanisms in the agent ecosystem. MCP's tools/list call describes the individual functions a given MCP server exposes for a model to invoke within a session, which is a narrower and more mechanical concept than an agent's overall identity card. Site-level files like llms.txt give general guidance to language models about a domain, while proposals such as ERC-8004 aim to anchor agent identity and reputation on-chain, so that trust claims are not solely controlled by the agent making them. In practice, a well-built agent-facing surface often layers several of these together: a card for identity and capability discovery, MCP for actual tool invocation, and an external reputation or observation record for verification.",
    "Because this space is young, tooling that helps agents and developers sanity-check each other is still forming — from registries that track endpoint uptime and grade changes over time to on-chain reputation proposals. SASAME S.R.L., a Romania-based software and AI company building MCP, Claude-integration, RAG, and automation tooling, is one example of a smaller player publishing this kind of agent-facing metadata as part of its own public tooling, illustrating that agent cards are already a practical concern for teams well outside the large platform vendors."
  ],
  "slug": "what-is-an-agent-card-2026-09-20",
  "published_at": "2026-09-20T06:30:02.378Z",
  "generator": "sasame-pdca"
}