The "Agent Host" Protocol: A Simple Guide to Microsoft's New Specification for Hosting AI Agents

Microsoft has published a new open specification called the Agent Host Protocol (AHP). Its goal is to standardize how applications interact with "AI agents," software assistants capable of carrying out tasks with a degree of independence, such as writing code or running commands in a terminal (the text-based interface used to control a computer). For readers with no technical background, this article explains the idea from the ground up.

What Is a "Host"?

In this context, a "host" is the system or program that runs and oversees an AI agent, much like a hotel hosts its guests. According to available sources, what is called an "Agent Host" is the subsystem responsible for hosting AI-powered coding agents, such as Microsoft's own coding assistant Copilot, and running them through the AHP protocol.

The central idea is that the host acts as the "source of truth," meaning it is the single authority that keeps the official, accurate record of a "session," which is an ongoing conversation or work task between a user and the AI agent. The applications people actually use, such as a code editor, are called "clients." These clients simply connect to the host in order to display and follow that session's status.

The Specific Problem the Protocol Solves

According to the available material, AHP addresses a specific problem: how can multiple applications (clients) connect to a single AI agent's work session and see the same status at the same time, instead of having each conversation and its pending tasks tied to just one application? As the sources describe it: "The conversation, its turn-taking, and any pending approvals for using certain tools were traditionally bound to a single application. If the user moved to another app or another device, they could not resume that same live session." AHP solves this by turning the session into a "shared resource" that can be accessed from anywhere, by any number of applications, at the same time.

One of the core principles behind this model is that an AI agent can keep working even if the application that started the session is closed. In other words, sessions are not tied to the lifespan of an application window; a window can be closed and the session later reopened from an entirely different window or device.

Who Is Behind the Specification, and Where Was It Published?

Microsoft published this open specification on its own GitHub platform. Its release coincided with a new version of the popular code editor Visual Studio Code, version 1.133, issued on August 12, 2026. This version introduced a separate process called the "Agent Host," which separates AI agents from the editor itself, enabling shared work sessions across multiple windows and remote execution, alongside the open AHP specification.

Step by Step: How Does the Protocol Actually Work?

The AHP workflow can be summarized in simplified steps. A "client" (the application) connects to the host and first receives an "initial state snapshot," a complete picture of the session's status at that moment. After that, it receives a stream of "ordered actions," which are the changes happening in the session as they occur, such as a new message or the result of an executed command. If the connection to the host is interrupted for any reason, the client can reconnect and receive either the actions it missed or a brand-new, complete state snapshot.

To organize this exchange, every element in the system is assigned a unique "address" or "channel," identified by a code beginning with "ahp," such as the root directory (ahp-root://), each individual work session (ahp-session), each conversation (ahp-chat), each terminal window (ahp-terminal), and each set of code changes (ahp-changeset). The client sends commands to the host, which responds with actions that update the state displayed to the client.

What Does the Protocol Deliberately Not Cover?

Sources make clear that AHP does not get involved in how the AI agent itself processes requests, how it calls external tools, or how it manages its "context window," meaning the amount of information the model can "remember" during a conversation. These matters remain the responsibility of the agent itself, and some may overlap with a separate protocol known as the Agent Client Protocol (ACP).

Likewise, AHP is not a substitute for the Model Context Protocol (MCP), another standard used by AI applications to connect to external tools and data sources. According to the available material, AHP is concerned exclusively with the "coordination layer" between clients and sessions, and it can work alongside both ACP and MCP without replacing either.

Where Can a Beginner Read More?

Readers who want to go deeper are advised to consult the official specification directly at this, where more precise technical details about the channels and actions described above are available. It is worth noting that most current sources on this topic are in English, and no Arabic-language or Middle East/North Africa coverage of this new protocol has been identified so far.