GitHub This piece argues that AI systems are settling into two distinct protocol layers. MCP standardizes how an agent reaches tools and systems. A2A standardizes how agents expose capabilities and collaborate with each other.
The architectural payoff is separation of concerns. Tool access and agent interoperability stop being coupled to a single SDK or runtime, which makes systems easier to scale, govern, and evolve.
Why the model is useful
- MCP works as the vertical rail: files, databases, APIs, and operational tools.
- A2A works as the horizontal rail: discovery, task assignment, streaming, and delegation across agents.
- Protocol boundaries reduce bespoke integration code inside orchestrators.
- Production-grade runtimes can be wrapped and reused instead of rebuilt from scratch.
What stands out
A practical insight in the article is that not every team needs to build agent intelligence from first principles. In many cases the right move is to wrap mature runtimes such as Copilot or OpenCode and make them composable through a standard interface.
That framing reflects the same engineering instinct seen in service meshes and API gateways: standardize the boundary so the ecosystem can grow without rewriting every consumer.
SDKs enable agent execution. Protocols enable agent ecosystems.