What happens when an AI agent needs help from another specialized agent?
In our previous demos, we showcased two separate agents:
- Buy Now Pay Later Agent, specialized in BNPL
- Card Management Agent, specialized in cards and transactions
Each agent is a complete application with its own specialized MCP tools. The agent doesn’t need to understand how those tools work. It simply knows how to use the capabilities available to it.
But what happens when a task crosses domain boundaries?
For example: “Show me my recent card transactions and tell me which of them were Buy Now Pay Later purchases.”
The Card Management Agent understands cards and transactions. It doesn’t need to know how the BNPL system works. This is where A2A (Agent2Agent) becomes interesting. In our latest demo, we enabled the Card Management Agent to collaborate with the BNPL Agent using A2A.
The architecture looks like this: User → Card Management Agent → BNPL Agent
The Card Management Agent continues to use its own MCP tools for card related tasks. When the user asks something related to BNPL, the Card Management Agent delegates that part of the task to the specialized BNPL Agent.
It doesn’t need to know:
- Which MCP tools the BNPL Agent uses
- How the BNPL backend works
- How BNPL data is retrieved
- How the BNPL Agent handles the task internally
That capability is encapsulated within the BNPL Agent. The Card Management Agent simply recognizes: “This is a task another specialized agent can handle.” And the BNPL Agent takes care of the rest.
This creates a clear separation of responsibilities:
- MCP connects an agent to its tools.
- A2A connects agents to each other.
MCP provides access to specialized capabilities, while A2A enables agents to collaborate without exposing their internal implementation.
This allows us to build specialized, domain specific applications instead of one giant agent that needs to understand everything.
In this demo, we’re taking the Card Management Agent from our previous video and giving it one new capability: It can ask another agent for help. Neither agent needs to know how the other works internally.
And that’s what I find most interesting about combining MCP + A2A.