Microsoft Calls Its AI Stack ‘Model-Forward, Agent-First’

Microsoft CEO Satya Nadella wrote in his annual letter to shareholders, published on 21 October 2026: “The tech stack we are building is model-forward and agent-first.” He added that “Models will continue to become more capable, while agents will take increasingly complex workflows through to completion.”

The letter then walks through the stack layer by layer, with a number at almost every step, from 88 new datacenters to 30 million paid Copilot seats. We have covered the same letter for its harness argument and its cost-to-outcome idea; this piece maps the stack itself.

Business Pill · MODEL ROUTING

A short Business Engineer explainer on model routing: sending each request to the model that fits it on cost and capability. It teaches the general idea only and says nothing about Microsoft or any company in this piece.

The Stack, Layer by Layer

The table follows the letter from the bottom of the stack to the top. The layer names are our labels; the descriptions and numbers are the letter’s own.

What Model-Forward, Agent-First Means

In plain terms, a model is the AI engine that reads and writes, and an agent is software that uses a model to carry out a task from start to finish. “Model-forward” describes a stack built to take in better models as they arrive; “agent-first” describes products designed around agents doing the work, rather than people clicking through apps.

The letter puts it this way: “With agents, we can now macro delegate and micro steer.” It says Microsoft 365 Copilot moved this year “from synchronous chat assistants to asynchronous agents that can complete complex, long-running tasks.”

The letter names what it calls the company’s differentiator: “This combination of widely used first-party agents, a multimodel system grounded in rich enterprise context, and hyperscale infrastructure is our differentiator.” It also says the stack is “open, interoperable, model-diverse, and built for heterogeneity at every layer.”

On money, it says Microsoft supports “business models that span subscriptions, consumption, advertising, and hardware.” Our page on Microsoft’s business model sets out how those revenue lines have worked so far, and our page on Microsoft’s mission statement covers how the company describes its purpose.

How Others Positioned on 20 and 21 October

Other AI companies set out their own positions on 20 and 21 October. The table is our framing of where each one put its weight, taken from our earlier pieces, which quote the companies or the outlets that reported them.

How We Read It

As we read it, the letter’s stack is unusual in how many layers carry a Microsoft product: its own datacenters and chips, its own MAI models next to third-party ones, its own context and control layer, and its own agents and apps.

As we read it, the others in the table each put more of their weight on fewer layers, as our earlier pieces describe them: Meta on a consumer agent and ads, Anthropic on its own app, OpenAI on distribution through other clouds, and Google on its own chips.

The Structural Read

As we read it, the letter’s stack has a Microsoft product at each layer it lists: datacenters, Maia and Cobalt chips, MAI models, Foundry and Work IQ, and Copilot.

As we read it, “model-forward” pairs with the catalog of “more than 11,000 models”: the stack is built to take in other companies’ models as well as Microsoft’s own.

As we read it, the positions others took on 20 and 21 October, as our earlier pieces describe them, each put more weight on fewer layers.

Three Implications

FOR ENTERPRISE BUYERS The letter says the stack is “open, interoperable, model-diverse, and built for heterogeneity at every layer.” As we read it, that is a claim buyers can test in contracts.

FOR RIVAL PLATFORMS As we read it, the second table shows the choice rivals face: compete across the stack, as the letter describes Microsoft doing, or own one or two layers well.

FOR INVESTORS The letter gives figures by layer but no revenue or profit by layer. As we read it, that split is the disclosure to look for.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Nine Layers of AI.

The framework’s starting point: the AI economy “is a layered industrial stack — and at any given moment, the binding constraint sits at a different layer than it did six months ago.”

As we read it, Nadella’s letter is Microsoft’s own account of that stack, with a product at each layer it names. The framework was published in May 2026 and does not discuss the letter.

What Is Not Established

The letter gives no revenue or profit figure by layer, and no figure for how much of its fleet runs on its own chips. It gives no count of customers that use agents at scale.

The second table is our reading of one or two days of announcements; it is not a full map of what each company sells, and the figures in it measure different things.

Connecting the Coverage

Our harness piece covers the same letter’s claim that models become interchangeable, and our cost-to-outcome piece covers its idea that customers will weigh token cost against outcomes. On 21 October we also covered Meta’s Muse in our Muse piece, Claude Duet in our Claude Duet piece and OpenAI’s models on Google Cloud in our Google Cloud piece. Google’s TorchTPU release of 20 October is in our TorchTPU piece.

What’s new here: the letter’s own layer-by-layer account of Microsoft’s stack, with the figures it gives for each layer, set against the positions others took the same week.

The Bottom Line

Nadella’s letter describes a stack that is “model-forward and agent-first,” with Microsoft products from datacenters and chips up to Copilot. As we read it, that sets Microsoft apart from rivals whose 20 and 21 October announcements centred on one or two layers.