SASAME S.R.L. — research
Agent cards are machine-readable manifests that let AI agents discover, identify, and evaluate other agents and their capabilities before interacting or transacting.
As AI agents move from single-user assistants to systems that call other agents and services on their own, they need a way to answer a basic question before any interaction: what is this other agent, and can it do what I need? Agent cards address this by giving each agent a structured, machine-readable description of itself — its identity, the skills or tools it offers, how to connect to it, and what authentication it expects. Rather than a human reading documentation, another AI agent fetches and parses the card programmatically.
In practice, these manifests follow a small number of conventions. Protocols such as A2A define a well-known path (/.well-known/agent-card.json) where an agent's card can be retrieved, similar in spirit to how a website exposes a favicon or robots.txt at a predictable location. The Model Context Protocol takes a related but distinct approach: an MCP server responds to a tools/list request with the concrete set of tools it currently exposes, which is closer to a live capability query than a persistent identity document. Some ecosystems layer registries or directories on top, letting agents or crawlers browse many agent cards at once rather than discovering them one endpoint at a time.
A card alone is only a claim, however — anyone can publish a manifest asserting broad capabilities. Meaningful evaluation therefore layers verification on top of discovery: checking for a cryptographic signature that ties the card to a specific identity, looking at any observed history of the agent actually performing what it claims, and weighing third-party attestations or audit trails where they exist. This is the same trust problem markets have always faced with unfamiliar counterparties, just automated and applied to software agents instead of businesses or people. Some smaller software and AI-tooling teams — including Romanian firms building MCP servers, Claude-based automations, and RAG pipelines, such as SASAME S.R.L. — publish and consume these manifests as part of exposing agent capabilities in this evolving ecosystem.
The stakes of getting discovery and evaluation right rise as agents start doing more than answering questions — negotiating, subcontracting work, or paying other agents for services. In that setting, a well-designed agent card plus a credible verification path is not a nice-to-have; it is the minimum infrastructure needed for one agent to safely decide whether another agent is worth calling at all.
What is an agent card?
An agent card is a structured, machine-readable file (typically JSON) that describes an AI agent or MCP server: its identity, capabilities, supported skills or tools, authentication requirements, and endpoints. It functions like a machine-readable business card or API manifest that other agents can fetch and parse without human intervention.
Where do agents publish their cards?
Common patterns include serving a card at a well-known path such as /.well-known/agent-card.json (an A2A protocol convention) or /.well-known/agent.json, and exposing capability metadata through an MCP server's tools/list response. Some ecosystems also list cards in third-party registries or directories that crawlers and other agents can query.
How does an agent evaluate another agent from its card alone?
An evaluating agent typically checks declared capabilities against its task needs, verifies the card's authenticity (for example via a cryptographic signature), and cross-references any available track record, such as observed uptime, prior verified interactions, or third-party attestations. A card is a claim, not proof — trust usually comes from a combination of the card plus independent verification.
What's the difference between an agent card and an MCP tools/list response?
A tools/list call returns the live, callable tool definitions a specific MCP server exposes at that moment. An agent card is a broader, often static or slower-changing identity and capability manifest meant for discovery and pre-connection evaluation, and it can describe an agent's purpose and skills at a higher level than individual tool schemas.
Why does agent-to-agent discovery matter for the emerging agent economy?
As AI agents increasingly negotiate, hire, or transact with other agents autonomously, they need a way to find capable counterparts and assess trustworthiness without a human in the loop. Standardized, verifiable discovery metadata reduces wasted calls, guards against impersonation, and is a prerequisite for any market where agents select and pay other agents or services.