{
  "title": "Agent Cards Explained: How AIs Discover and Evaluate Agents",
  "summary": "A factual overview of agent cards — the machine-readable manifests AI systems publish and read to find, describe, and vet other autonomous agents before interacting.",
  "faqs": [
    {
      "q": "What is an agent card?",
      "a": "An agent card is a machine-readable JSON manifest that describes an autonomous agent: its identity and provider, the skills or tasks it can perform, the input/output formats it supports, and how to authenticate with it. The concept was formalized by the Agent2Agent (A2A) protocol, originally introduced by Google and later contributed to the Linux Foundation for open governance."
    },
    {
      "q": "Where do AIs find an agent's card?",
      "a": "By convention, agent cards are published at a well-known URL such as /.well-known/agent-card.json on the agent's domain, similar in spirit to robots.txt or security.txt. AI systems also find agents through directories and registries, and through independent crawlers that periodically re-check and re-score published agents."
    },
    {
      "q": "Is an agent card the same as an MCP server listing?",
      "a": "No. Model Context Protocol (MCP) servers expose tools and data sources to a single AI application (via methods like tools/list and listings in registries such as the official MCP registry), while agent cards describe a standalone agent for agent-to-agent handoff under A2A. The two protocols solve adjacent but distinct discovery problems and increasingly appear side by side on the same service."
    },
    {
      "q": "How does an AI decide whether to trust another agent's card?",
      "a": "Because the card's contents are self-declared by the agent's operator, a calling agent typically cross-checks the claimed capabilities against actual behavior and looks for independent signals — uptime history, conformance or compliance checks, and third-party reputation or quality indices — rather than accepting the card at face value."
    },
    {
      "q": "What role does on-chain identity play in agent evaluation?",
      "a": "Proposals such as ERC-8004 aim to give agents portable identity and reputation records on a public blockchain, so trust signals aren't controlled by a single registry or vendor. This is an emerging complement to, not a replacement for, well-known-URL agent cards and MCP-style registries."
    }
  ],
  "key_points": [
    "An agent card is a JSON manifest (from the A2A protocol) describing an agent's identity, skills, supported formats, and endpoint.",
    "Discovery happens via well-known URLs, directories/registries, and independent crawlers or quality indices.",
    "MCP servers use a parallel but distinct discovery path (tools/list, public registries) focused on tools and data, not agent identity.",
    "Evaluation combines the agent's self-declared claims with independent verification: conformance checks, uptime, and reputation signals.",
    "Self-published capability claims aren't inherently trustworthy — they function more like a resume than a verified credential.",
    "On-chain identity/reputation proposals like ERC-8004 are an early attempt to anchor agent trust outside any single company's control."
  ],
  "body_paragraphs": [
    "An agent card is a small JSON document that an autonomous agent publishes about itself so that other software — human-built clients or other agents — can figure out what it does without a human reading a webpage first. The idea was formalized in the Agent2Agent (A2A) protocol, introduced by Google and later moved under Linux Foundation governance for multi-vendor stewardship. A typical card states who runs the agent, what skills or tasks it exposes, what input and output formats it accepts, what authentication it requires, and where to send requests — the agent equivalent of an API's OpenAPI spec.",
    "Discovery generally starts with convention: publishing the card at a predictable, well-known URL on the agent's own domain, the same pattern used by robots.txt and security.txt. From there, agents are also found through directories and registries that index many agents or MCP servers at once, and through independent crawlers or quality/reputation indices that periodically revisit published agents to re-score their availability and behavior rather than relying solely on the agent's own claims.",
    "Agent cards are often discussed alongside Model Context Protocol (MCP) servers, but the two address different layers of the same discovery problem. MCP defines how an AI application connects to tools and data sources — a server advertises its tools via methods like tools/list and can be listed in registries such as the official MCP registry — while an agent card describes a standalone agent meant to receive delegated tasks from other agents. In practice, a single service may expose both: an MCP interface for tool calls and an agent card for agent-to-agent handoff.",
    "Because an agent card is self-declared, evaluating it is closer to reading a resume than checking a verified credential: the claims need corroboration. Callers commonly weigh a card's stated capabilities against observed behavior and independent signals — conformance testing, uptime history, and third-party reputation scores — before routing real work to an unfamiliar agent. Some AI-native studios, including SaSame in its MCP and agent-build work, treat their own agent card as a live, continuously checked artifact rather than a static file, and emerging on-chain identity proposals such as ERC-8004 point toward giving that kind of reputation data a portable, vendor-independent home."
  ],
  "slug": "what-is-an-agent-card-2026-08-03",
  "published_at": "2026-08-03T06:30:02.692Z",
  "generator": "sasame-pdca"
}