Protocols for building LLM centric agentic workflows
Summary
This post outlines the main communication protocols agents and agentic workflows are using for communicating with tools, applications or between the agents. I have a separate page which covers a range of frameworks and toolsets used to building Agentic AI applications.
Below listed communication protocols define how agents and agentic, LLM centric workflow applications integrate with various data sources, tools and applications.
Protocols
- Model Context Protocol (MCP)
- Universal Tool Calling Protocol (UTCP)
- Agent2Agent Protocol (A2A)
- Agent Communication Protocol (ACP)
- Secure Low-Latency Interactive Messaging (SLIM)
MCP
Model Context Protocol or MCP is an open source protocol which aims to standardize how AI applications connect and interact with external systems.
Originally started as an open source protocol project by Anthropic - its now part of Agentic AI Foundation goverened by a wider Linux Foundation.
MCP bacame industry satandard protocol for how agents integrate and use tools, its being adopted by a growing group of popular application and service providers.
MCP defines a MCP client application (agent like Claude Code, Claude Desktop, Cursor and others) and MCP server - application which translates agent instructions to the specific formats and protocols external tools can understand.
MCP servers also privide a list of available tools to agents (including tool descriptions) and agents themselves decide when and how to use the tools.
UTCP
Universal Tool Calling Protocol (UTCP) is a protocol which only enables discovery of descriptions for available tools.
It takes a different route than MCP - agents see the connection details for tools and services and communicate with them directly, without middle layer mandated by the MCP protocol.
For example, agent will discover the connection details and methods for the specific API service endpoints and will make requests to it directly.
UTCP serves as a description protocol and not placing itself in the middle of agent requests. It's open source specification.
A2A
Agent2Agent or A2A protocol is an open source protocol for agentic workflows, which defines the standard for agent to agent data exchanges.
Its aim to complement MCP protocol with the focus on the multi agent systems, where agents interact with each other.
Proposed, developed and maintained by Google.
ACP
Agent Communication Protocol or ACP is an open protocol for communication between AI agents. As an alternative to Agent2Agent protocol listed above.
SLIM
SLIM - Secure Low-Latency Interactive Messaging is a messaging framework that provides the transport layer for agent communication protocols like A2A. While A2A defines what agents say - message formats, semantics, coordination patterns, SLIM defines how these messages are delivered across networks.
Notes
New protocols are being discussed and rolled out in the industry. Let me know if I'm missing some on this list and if you would recommend additional protocols to be added to the list.