The AX Stack: Which Agent Design Layers Are Settled and Where Opportunity Remains

The term "AX Stack"—Agent Experience Stack—has been circulating in private engineering conversations for months, but Microsoft has just given it a public name and structure in a developer blog post. The article, titled The AX Stack: What's Fixed & Where You Can Win, proposes a framework for mapping which layers of agent development are sufficiently consolidated to build on without the ground shifting, and which remain territory where teams can differentiate—or where you simply have to accept uncertainty.

For those of us working daily with Claude Code, MCP servers, and sub-agents, the distinction is far from trivial: investing weeks in a layer that will change specification next quarter carries real cost.

What Microsoft Means by "Fixed" and "Open"

The article divides the stack into two broad zones. The fixed zone groups the pieces where there is enough convergence in the industry to assume reasonable stability: tool-calling protocols, structured message formats, and the foundations of reasoning in large language models. This block includes things like MCP, which Anthropic has developed as an open standard with implementations already across multiple environments—Claude Desktop, Claude Code, third-party integrations.

The open zone is where the article becomes most interesting: agent evaluation, long-term memory management, coordination between sub-agents, and especially everything related to the user experience when someone interacts with an agentic system. There is no dominant standard here, and the space to build something differentiated is real.

Why Layer Separation Matters

A frequent mistake when designing agent pipelines is treating the system as a monolith. When the model, the tool protocol, orchestration logic, and user interface are tightly coupled, any change in one layer forces you to rewrite the others. The AX Stack proposal pushes in the opposite direction: identifying clear contracts between layers so each can evolve independently.

In practice, for a team using Claude Opus 4.7 with a one-million-token context window and several specialized sub-agents orchestrated from Claude Code, this translates into concrete questions: should the routing logic between sub-agents live in a `PreToolUse` hook? Or in a skill dedicated to coordination? Does shared state between agents sit in context memory or in an MCP server acting as a persistence layer? The AX Stack framework doesn't answer these questions directly, but it does help you formulate them with greater precision.

Who This Analysis Is For

The article is written with product teams and architects in mind—people making platform decisions—not developers looking for code samples. If your job is deciding what to build internally and what to buy or reuse from the community, the proposed taxonomy has immediate value.

For integrators working specifically with the Claude ecosystem, there is an additional reading: the fact that MCP is in the "fixed" zone according to Microsoft—a company with its own competing agentic stack—is a convergence signal worth paying attention to. When two major players with different interests agree that a protocol is infrastructure rather than competitive advantage, that is usually a good sign to bet on it.

What the Article Doesn't Cover

Microsoft's analysis focuses on the developer side and barely touches the end-user experience: how to communicate to a person what an agent is doing on their behalf, how to manage consent for high-impact operations, or how to design interfaces that don't create anxiety when an agent makes autonomous decisions. These are, paradoxically, some of the most "open" layers of the stack and also the ones that will most determine whether agentic systems gain real adoption outside purely technical environments.

The Hacker News thread for the article had generated no comments at the time of publishing this post, which says something about how specialized the audience still is for this type of structural discussion.

---

We value Microsoft naming the layers of the agentic stack: the industry has been building on sand for months without a map. The AX Stack is not that definitive map, but it is a useful starting point for teams that need to decide where to concentrate their own effort.