In traditional software systems, REST has been the go-to standard for structured communication between services. But as we begin to build agentic AI systems—where autonomous agents plan, act, and collaborate—the REST model starts to show its limits.

Recently, I’ve been working with Message Chain Protocol (MCP), a protocol designed specifically for agentic systems. It’s made me think: in the same way REST shaped web APIs, could MCP become the backbone for multi-agent orchestration?

What Is MCP?

MCP (Message Chain Protocol) is a messaging format designed for AI agents to collaborate and complete tasks. Instead of one service calling another (like in REST), MCP works more like a relay baton—a message moves through a chain of agents, each adding value before passing it on.

At a high level, an MCP message includes:

• A task or goal
• State or context (like partial progress)
• A chain of agents or steps to follow

It’s asynchronous, decentralized, and built for autonomous workflows—making it well-suited for multi-agent AI systems where tasks evolve and adapt along the way.

Why REST Doesn’t Quite Fit Agentic AI

REST was built for stateless, client-server interactions. You request a resource, you get a response. That model works well for CRUD operations, but starts to feel rigid when agents:

• Need to initiate autonomous tasks
• Operate asynchronously
• Chain tasks together with feedback loops
• Make decisions dynamically as new information arrives

Imagine a team of AI agents handling a research task. One agent drafts, another edits, another fact-checks. Trying to coordinate that over REST quickly leads to tangled logic and brittle control flows.

What Makes MCP a Better Fit

Autonomous Flow Control

Each agent can decide when and how to act based on message content—not because it was directly called. This fits naturally with agentic AI where planning is often emergent or adaptive.

Feedback Loops and Self-Correction

Agents can append new handlers to the chain based on results. This enables self-refinement, retries, or fallback strategies—difficult to express in a RESTful design.

Loose Coupling and Flexibility

MCP abstracts communication from transport. Agents don’t need to know about URLs, endpoints, or protocols—they just handle messages. This makes the system more modular and easier to evolve.

Asynchronous by Default

MCP is naturally async, which is critical when agents run long tasks (e.g. querying web data, writing documents) or work in parallel. REST, by contrast, assumes sync blocking behavior unless you bolt on pub/sub patterns.

So, Is MCP the “New REST”?

In the context of agentic AI systems, I’d say yes—it could be.

Not in the sense that it will replace REST for web APIs, but in the way that REST became the lingua franca of the web, MCP might become the shared protocol for AI agents.

We’re already seeing this pattern emerge in advanced orchestration layers—like LangGraph, Autogen, and agent frameworks that require multi-step planning and task handoff. Many are converging toward some variant of message-based workflows.

A Hybrid Future?

Just like we didn’t throw out REST when GraphQL or gRPC arrived, I don’t expect MCP to “replace” REST. But it’s likely to augment it, especially behind the scenes in systems where agents, tools, and autonomous workflows need to interoperate smoothly.

My current approach? Use REST at the edges—for external integrations or UI interactions—but use MCP internally to let agents plan, execute, and collaborate.

Final Thoughts

If you're building agent-based applications or orchestrating complex AI workflows, MCP is worth serious consideration. It reflects the way agents actually operate: message-driven, adaptive, and decentralised.

We’re still early in this space—but it feels a lot like the early days of REST. That’s exciting.