How do AI Agents work together? How do we get agents from our company working with agents from the main AI tools?
AI agents are programs designed to perform specific tasks. Agents often struggle to work together because they are built by different companies using different protocols and systems.
The A2A (Agent-to-Agent) Protocol is an open standard that helps to address this by acting as a universal translator. The focus is on helping AI agents communicate and collaborate to complete a single task.
One of the main components of this protocol is an Agent Card. An Agent Card provides basic information about the agent, lists what capabilities and skills the agent has, and details what authentication systems should be used so that the agents can securely work together.
Without these cards, an “Inventory Agent” and a “Shipping Agent” from different companies wouldn’t know how to talk to each other. With an Agent Card, they can “discover” each other and collaborate automatically.
Technically, an Agent Card is a small machine-readable file written in JSON. It is usually hosted in the /.well-known/ directory with the file name agent.json. This standard location makes it easier to request and retrieve the AgentCard. (Same idea for why the robots.txt file is hosted in a standard location.)
The AgentCard structure typically contains the following information:
- Identity & Metadata: This includes the agent’s name, URL, a brief description of what it does, its version number, and a link to its icon.
- Capabilities: This is a short list of true/false values for specific capabilities supported or not by the given agent.
Skills: This is a more detailed list of the specific tasks the agent can perform. For example, “book_flight” or “generate_invoice” or “contact_customer_support”. Each skill includes a name, description, and tags for categorization. Importantly, skills also include an area to provide example scenarios or prompts where this agent can help.
- Input and Output Modes: This defines how the agent sends/receives information and what types of files it can send/receive.
- Authentication Requirements: This provides security details and specifies credentials.
So, why does this matter? If the future is agentic AI, and right now that is a pretty big if, then every company will likely need to produce AI agents that are capable of working with other AI agents.
If that is the future, then an AgentCard becomes an important tool to promote a company. Optimizing the agent’s description or skill information will help other agents know how and when to work with your company’s agent to complete a task. Without a properly configured and optimized AgentCard, your company could lose business.
That’s all future speculation. Right now is the time to experiment. If you do have a service that you can turn into an AI Agent, using an AgentCard would be a method to use to start testing how well that card works to help other agents connect with it.