Why Agent-to-Agent Communication

Agents are not just tools

A tool is a function: fixed inputs, an immediate answer. A remote agent reasons, may need to ask follow-up questions, works for minutes or days, and uses its own tools and model. Wrapping a whole agent as a tool call hides this: you lose status updates, clarifications, cancellation and partial results. A2A (Agent2Agent) is an open protocol, developed in the open (originating at Google and now hosted by the Linux Foundation), for agents built by different teams, vendors and frameworks to discover each other, delegate tasks, stream progress and exchange results, without sharing internal memory, prompts or tools (the remote agent is "opaque"). Specification details evolve, so verify against the current official documentation.

Discover, delegate, track

A2A lets independent agents advertise skills in an Agent Card, delegate tasks and exchange results.

Tool call versus agent task

Why long, interactive work needs more than a function call.

Tool call (MCP)                         Agent task (A2A)
fixed input -> immediate output          goal -> may take minutes/days
no memory of its own                     has its own reasoning, model, tools
errors are final                         can ask for input ("which invoice?")
no progress updates                      status updates, streaming, artifacts
you see the tool's inputs/outputs        you see only what the agent chooses to share (opaque)

Quick check: Why is "opaque" a useful property of A2A remote agents?

• It makes agents invisible to users
• Agents can collaborate without exposing their internals
• It removes authentication
• It speeds up GPUs