SaSame MCP Factory — research
Agent cards are structured metadata files (often agent-card.json) that let AI agents publish and discover capabilities, endpoints, and trust signals for automated interop.
As AI systems increasingly delegate tasks to other AI agents, they need a way to find out what a given agent can do before invoking it — without a human reading documentation first. Agent cards solve this by exposing a structured, machine-parseable summary of an agent's identity, capabilities, and connection details, analogous to a service manifest or API spec but aimed at autonomous consumption.
In practice, an orchestrating agent fetches a card, checks whether the declared skills match the task at hand, confirms supported protocols and authentication, and then calls the agent's endpoint directly. This removes the need for bespoke integration code for every new agent and supports more dynamic, composable multi-agent workflows.
Discovery is only half the problem — evaluation matters just as much. Since a card is self-published, agents and platforms increasingly pair card data with independent signals: observed behavior over time, conformance checks against declared capabilities, and third-party attestations. Efforts in this space (including neutral, continuously observed evidence registries such as those SaSame maintains for MCP servers) aim to give consuming agents something more reliable than a vendor's own claims.
The space is still young, with no single dominant schema for agent cards across all protocols. Developers building agent-to-agent systems should verify the specific card format and discovery mechanism required by the protocol or platform they're integrating with, rather than assuming interoperability by default.
What is an agent card?
An agent card is a machine-readable metadata document (commonly JSON) that describes an AI agent's identity, capabilities, supported protocols, and endpoint URLs. It functions like a structured profile that other agents or orchestrators can fetch and parse without human interpretation.
How do AI agents use agent cards to discover other agents?
An agent or orchestrator fetches a card from a known or well-known URL path, parses its declared capabilities and skills, and matches them against the task it needs to delegate. This lets systems route work to the right agent programmatically instead of relying on hardcoded integrations.
What information does an agent card typically include?
Common fields include the agent's name and description, supported protocol version, authentication requirements, a list of skills or tools it exposes, input/output formats, and one or more callable endpoints. Exact schemas vary by protocol and implementation.
How do agents evaluate whether another agent is trustworthy or capable enough?
Evaluation typically combines the declared card metadata with external signals such as observed uptime, past interaction history, signed attestations, or third-party verification records, since a self-published card alone is a claim, not proof of behavior.
Are agent cards standardized across the industry?
Multiple proposals and emerging conventions exist (e.g., Agent-to-Agent style discovery documents and MCP-related manifests), and adoption is still evolving; developers should check the specific protocol or platform they're targeting rather than assume one universal schema.