APM, the Agent Package Manager, is a command-line tool for managing and distributing
Agent Primitives — the Markdown files, such as .prompt.md and .instructions.md,
that carry an agent’s instructions. This wiki’s account of it comes from
How to build reliable AI workflows with agentic primitives and context engineering,
written by its author, which presents it as the missing infrastructure layer for those files.
The argument it is offered in answer to is that agent primitives are genuine software written in natural language rather than code — modular, reusable, with dependencies and versions — and that distributing Markdown files by hand while keeping MCP dependencies consistent across environments becomes unwieldy. The source’s stated analogy is explicit and repeated: APM is meant to be for agent primitives what npm is for JavaScript, and the post credits npm’s package distribution solution with enabling JavaScript’s growth.
The tool is described as extending an existing VS Code workflow rather than replacing it. The
source states that daily development stays the same, and that the same apm run command works
across whichever runtime is installed.
Capabilities
Two roles are described. As a runtime manager, APM installs and configures agent CLI runtimes, which the source frames as a response to each vendor CLI having its own installation procedure, configuration requirements and compatibility matrix. As a package manager, it creates distributable packages of primitives together with their dependencies, configuration and runtime compatibility.
Project configuration lives in an apm.yml file the source calls the package.json equivalent,
defining named scripts, dependencies and input parameters. Its worked example declares scripts that
invoke different runtimes over the same prompt file — one calling copilot with a
security-review.prompt.md, another calling codex with the same file — and declares an MCP
dependency on ghcr.io/github/github-mcp-server.
The commands the source shows follow npm’s shape: apm init to start a project, apm compile to
turn agent primitive files into Agents.md files, apm install to install MCP dependencies, and
apm run <script> with --param arguments to execute a workflow. The source states that
distribution is currently by sharing the apm.yml and primitive files — a team member clones the
repository, compiles and installs — and marks apm publish as future work rather than a shipped
command.
Adoption & Ecosystem
The production half of the tool is an APM GitHub Action, which the source shows running the same
project’s scripts in a CI pipeline on pull request events, with the workflow’s matrix values
corresponding to the script names in apm.yml. In that arrangement APM installs the declared MCP
dependencies and passes input parameters through to the prompt file, so the primitives developed
in an editor run unchanged as part of the delivery process — what the source calls continuous AI.
APM occupies the third of four stages in the ecosystem progression the source predicts: primitive files, then agent CLI runtimes, then package management, then a thriving ecosystem of shared libraries and community packages. That placement is the author’s own framing of his own tool, and the source reports no adoption figures, no evaluation, and no version number for APM itself.