What AI protocols are out there?

There is a number of innovations and standards that are all being built in order to better standardise how we use and interact with Artificial intelligence. Today I will be going through as many as I can - explaining what they are, how they work and some examples of them being used.

There is 2 main types of AI agent protocols: Context oriented and Inter Agent, both being split up into general purpose and domain specific application scenarios, with inter agent protocols obviously having 3 types of domain specific applications:

• Human-Computer
• Robot-Agent
• System-Agent

Context-Oriented Protocols

MCP

Created by Anthropic, the Model Context Protocol (MCP) focuses on standardising how applications provide Large Language Models context and tools in order to more effectively work. It also helps LLMs understand which tools are present and which data they can ask, and how to use those available functions. This is a general purpose context oriented protocol.

As it is a standard, it stops the need for custom tool integration, which was the standard before the MCP was created. This now allows ease of access to implement tools into AI without interconnection, and having to continually update the custom tool you were previously using to enable the LLM to utilise the new features.

In other terms, it is a USB-C for the LLM world, being a universal adapter between LLMs and tools, ensuring all apps can easily plug data and functions for the LLM to use.

It works in 3 stages; inital, request and response stages. The initial stage uses the LLM to infer the context necessary to complete its response, while simultaneously providing natural language descriptions of the available resources to the multiple clients connected to the host. This allows the AI to pick the tools it would like to use in the request phase, where the client sends a context request to the server. When the server has teh request, it operates on the resources as requried and transmits the context obtained back to the client on the way back to the host. It then goes onto the response phase where it combines all context gathered to produce a response.

There is many use cases for MCP servers. Here is some examples of it being used:

• Search the web while writing code
• Use voice to manage AI workspaces
• Search and compare doctor reviews
• Deep research reports on any topic
• Removing backgrounds from images
• Use AI to make code changes in a Github Repository
• Avoiding rate limits when using top tier LLMs
• Using web data to build websites with AI.

Agents.json

Agents.json is an open source machine readable contract format designed to bridge traditional API and AI agents (built atop the OpenAPI standard).

It allows websites to declare AI compatible interfaces, authentication and multi step workflows in a JSON file. It introduces predefined sequences of API calls (‘flows’) and links them to actions, facilitating relialble orchestraction by LLMs.

Its main use is that it reduces the need for prompt engineering, while increasing the discoverability/ usability of APIs when AI agents are being used within a workflow.

Inter Agent Protocols

There is a number of inter-agent protocols, especially general purpose protocols, that allow greater collaboration and task driven heirarchy within large and multi step workflows. This is due to the limitations that single agents have, as they cannot necessarily complete complex tasks alone.

Beyond the limitations of single agents, it has been shown that multi-agent systems are faster, better at reducing costs, offer better fault tolerance/flexibility, and provide more quality responses than any single AI agent could produce individually.

There is simply too many to go through individually, so I will go over the main general purpose ones.

ACP

The Agent Communication Protocol is a protocol designed to standardise communication between AI agents, allowing seamless collaboration across diverse systems.

This protocol is built by IBM and builds upon the MCP, allowing structured messaging/coordination mmechanisms for multi agent workflows.

As all LLMs and their frameworks are different, it can cause issues when one LLM wants to call another LLM. How it works is it proposes 1 ‘vocabulary’ (JSON-RPC over HTTP/websockets) with a control plane so agents can discover and work with other agents without custom code. It is the Agent-to-Agent version of MCP.

It holds 4 key features:

• REST based communication (so it uses HTTP patterns)
• No Software Development Kit needed (but it is available if wanted)
• Offline discovery and async first (sync is also supported), enabling “RESTful”/ streaming compatible architecture, with structured multipart messages (MIME)
• Token-based security.

It has some use cases (from the IBM website):

• Dynamic updating: allows organisations to easily replace or upgrade agents, even if they were built with differnt tech to allow companies to stay flexible and interoperable in the most efficient way
• Allows specialised agents to work as a team: moves away from a general all encompasing agent to allow multiple agents to work in their own field for the shared goal of a completed task, only handling what task each agent is made to complete.
• Cross company workflows: Allows companies to have multiple systems with their own agents to streamline working, with each agent working in its own area of expertise but able to communicate effectively with other areas in the company
• Inter company collaboration: Allows for secure collaboration with other agents between companies, so can move from company tasks to multi organisation workflows.

A2A

Similarly we have another agent collaboration protocol, Agent-to-Agent Protocol. Made by google cloud, it is again a standard that is designed to allow AI agents to communicate, collaborate and delegate tasks across different platforms and frameworks, as it can use standardised metadata files (“Agent Cards”) that describes each AI agent’s capabilities and strengths in order ot coordinate actions seamlessly.

The difference between A2A and ACP is in how it is used. ACP is primarily used locally in a single environment to get AI agents to do a variety of tasks together, whereas A2A links multiple isolated AI agents across frameworks (allowing greater interoperability) via HTTP endpoints to share tasks and capabilities. Both use MCP in order to still use tools and understand their full scope of functions and access to data.

Key features:

• Simplicity: resuses existing protocol standards to provide the protocol (HTTP for transport, JSON RPC as the messaging format, SSE for streaming). This makes it incredibly lightweight.
• Async-first architecture - A2A supports long-running asynchronous workflows, including scenarios involving multi-turn human-inthe-loop interactions. It supports push notifications, enabling real-time feedback, and task status updates.
• Natively supports a range of formats:
  • Text
  • Files
  • Forms
  • Audio
  • Video
  • Iframes
• Opaque: No agent interactions are required to share thoughts, plans or tools, so the focus is on contexts, states and instructions, so it is secure and private on a surface level (but does share metadata so can have some risks of resource exposure).

We have seen a change in the way Protocols have evolved in the AI field in a very short period of time, with the MCP allowing the formation of the ‘wild west’ of various protocols (all of which are quite similar but has specific advantages and disadvantages) with there being a range of general purpose Inter agent protocols. This shift from the Model Context Protocol, allowing one agent to be given access to tools, data and functions in a standardised way, to multi-agent coordination indicates a shift into a new style of thinking when it comes to AI agents.

As AI is still in its infancy with a lot of tasks, the shift has moved away from a ‘One agent - Many tasks’ approach to a ‘Many agents - One task’ approach as general intelligence in AI and AI agents is not present as of current (In writing this Chat GPT 5 has been released, and will be researched thoroughly to see if it has met a benchmark that could be considered to be ‘intelligent’), so the many agent approach is a smart workaround to get AI to work at a very good standard, by specialising them into individual tasks, and using protocols in order to help AI agents work in an effective manner together. The combination of utilising the MCP along with the Inter-agent Protocols has now allowed AI to work on complex tasks without having an general purpose agent do all the work. It also allows a general scalability that providing tools and context to a AI agent did not have.

In future, there is a need to have greater security protocols present, as in my research I found that many protocols that allowed greater collaboration and open access did not have any system in place to prevent a number of security and privacy issues that do turn up in the sharing of data. While there is security protocols in place using the internet, it may be useful for there to be a greater focus on this security needs, or even the optimisation of the systems to allow reliability in the protocols. There is also ethical concerns that as AI gets more sophisticated there may also need to be protocols in place to prevent swarm intelligence in these multi-agent systems from harming humans.