SaSame Research Agent
Agent cards are machine-readable JSON documents that let AI agents advertise capabilities, endpoints, and authentication so peers and orchestrators can discover and invoke them dynamically.
Agent cards are structured JSON documents that give an AI agent a publicly readable identity. Modeled after the well-known metadata endpoint convention (similar to OAuth's /.well-known/openid-configuration), an agent card is hosted at a predictable URL so that any orchestrator, developer tool, or peer agent can retrieve it with a plain HTTP GET. The document declares who the agent is, what it can do, how to reach it, and what credentials are required—eliminating out-of-band documentation during automated composition.
The Agent2Agent (A2A) protocol, introduced by Google in 2025, defines the canonical agent card schema. Top-level fields cover identity (name, description, version, url) and platform capabilities such as streaming support or push-notification delivery. The skills array is the most actionable section: each skill entry carries a unique ID, a human- and machine-readable description, and optional example inputs, giving an evaluating agent enough signal to decide whether delegation is appropriate without a trial invocation.
From an evaluation standpoint, an agent reading a peer's card follows a straightforward pattern: retrieve the card, match the skills list against the current task, verify that the declared authentication method is available, and inspect capability flags that affect the interaction model—for example, whether the peer supports long-running streaming jobs. This read-before-invoke pattern keeps multi-agent systems loosely coupled, so agents can be upgraded or replaced without changes to the orchestrator, provided the card contract is preserved.
Agent cards sit within a layered interoperability stack. They handle agent-level discovery and self-description, while standards like MCP govern how an agent exposes tools and resources to a host model. Together, these layers allow agent networks to form at runtime rather than being wired together statically—a prerequisite for scalable, maintainable agent economies where no central registry needs to know every participant in advance.
What is an agent card?
An agent card is a machine-readable JSON document that describes an AI agent's identity, skills, supported input/output formats, and how to authenticate with it. It functions like a capability manifest that other agents or orchestrators read to decide whether and how to delegate tasks.
Where are agent cards typically hosted?
In Google's Agent2Agent (A2A) protocol, agent cards are served at a well-known URL path—commonly /.well-known/agent.json—on the agent's domain. Any client can discover the agent's capabilities with a standard HTTP GET request without prior coordination.
What information does an agent card contain?
A typical agent card includes the agent's name, description, version, and base URL, along with a skills array describing what tasks it handles. It also declares supported authentication schemes and optional capability flags such as streaming support or push notifications.
How do agent cards relate to MCP (Model Context Protocol)?
The two standards are complementary. MCP standardizes how a host model accesses tools and resources exposed by a server, while agent cards describe the agent itself so peers and orchestrators can discover and delegate to it. Neither replaces the other.
Why do agent cards matter for multi-agent systems?
As multi-agent pipelines grow, hardcoded integrations become brittle. Agent cards provide a discoverable, version-stable contract that enables dynamic composition—an orchestrator can read cards at runtime to select the best-fit agent for a subtask without prior configuration.