Microsoft Wants a Package Manager for Your AI Agents

Software developers solved a version of this problem a long time ago. When your code depends on someone else's code, you do not copy and paste it and hope for the best. You declare the dependency, pin it to a version, and let a tool fetch exactly what you asked for. Microsoft now thinks AI agents need the same discipline, and it has built an open-source tool called APM to provide it.

The pitch, in Microsoft's own words on the project site, is a dependency manager for AI agents that works like npm for agent context. If you have ever run a package manager, the shape of it will feel familiar. You declare what your agent needs, and apm install pulls it in, resolving the chain of dependencies behind it automatically.

What an agent depends on

An AI agent is more than a model. It is a bundle of instructions, skills, prompts, and configuration, the accumulated rules and examples that tell a general-purpose model how to behave for a specific job. Today that material tends to live in scattered files, pasted into repositories and copied between projects, with no reliable way to say which version you are running or where it came from.

APM treats those pieces as packages. One package can depend on another, and APM works out the full tree, the same way a web project pulls in a library that quietly pulls in a dozen more. For teams trying to get a coding agent to behave the same way on every developer's machine, that reproducibility is the whole point. A companion command, apm install --frozen, installs only what a lockfile specifies, which is the kind of guarantee a continuous-integration pipeline needs before it will trust a build.

Built to cross the fences

The more interesting choice is that APM does not tie itself to Microsoft's own products. The project, released under a permissive MIT license, aims to work across the major coding assistants, including GitHub Copilot, Cursor, Claude, Codex, and Gemini. In a market where every vendor would prefer to own the whole stack, a neutral format that moves an agent's configuration between rival tools is a small act of defiance.

It is early. The command line tool is still at an initial 0.8 release, and a package manager is only as useful as the ecosystem of packages around it, which for now is thin. There is also a real question of whether the industry will standardize on Microsoft's format or on one of the several competing attempts to do the same thing.

What APM captures, though, is a shift in how people are starting to treat agents. The novelty of getting a model to do something clever is wearing off, and the harder work of making it do that thing reliably, repeatedly, and in a way a colleague can reproduce is moving to the center. Tools like the ones being tested on real scientific research only matter if their behavior can be pinned down and shared. That is a less exciting story than a new model launch. It is also the sign of a field growing up.