A support agent might ask a scheduling agent to arrange an appointment. The scheduling agent might then use a messaging tool to send the confirmation.
Those interactions have different boundaries. One delegates work to another agent. The other invokes a capability exposed by a service.
What does A2A provide?
A2A provides an interface through which agents discover capabilities, exchange messages and manage work across separate systems.
The A2A specification defines an Agent Card that advertises an agent's capabilities and connection information. A caller can use that information without access to the remote agent's internal implementation.
An interaction can return a message directly or create a task. Tasks have identifiers and states, so the caller can follow work that continues beyond the first response. An artifact carries an output produced during that work.
For the appointment example, a scheduling agent could return a task while it waits for missing details. Its result could then identify the confirmed appointment.
What does MCP provide?
MCP provides a shared interface for an AI application to discover and use tools, resources and prompts.
The MCP specification separates the host application, its client connector and the server providing capabilities. A tool exposes an executable function. Resources supply context, and prompts supply reusable message templates.
In the appointment example, a messaging server could expose a tool that sends the confirmation. The scheduling agent's host invokes that tool with the recipient and message content.
An MCP tool can also wrap an agent. The distinction concerns the interface exposed to its caller, rather than whether a model exists behind it.
How do the interfaces compare?
A2A describes exchanges with a remote agent. MCP describes access to the capabilities a server exposes.
| Question | A2A | MCP |
|---|---|---|
| How does the caller discover the interface? | An Agent Card describes the agent | The server advertises capabilities and exposes listings |
| What does the caller submit? | A message or an operation on a task | A request such as calling a tool or reading a resource |
| What can come back? | A message, task state or artifact | A result defined by the requested operation |
Both specifications describe more than an immediate request and response. Long-running work alone does not decide which interface fits. Choose around the contract your caller needs to use.
How would I connect the messaging step to Bird?
You can connect your agent to Bird's MCP server to send a message and inspect its outcome.
Keep the identifier returned by the send. A later read reports the recorded delivery status, which is separate from the success of the initial request.
Your application can place this messaging step inside a larger agent workflow. The messaging step uses Bird's MCP interface.
For email, a delivered result reports acceptance by the receiving mail server. It does not establish that someone read the message. The delivery-state answer explains the channel-specific boundaries to use when reporting completion.