At Build 2026, Microsoft announced a new category of AI: autopilots. Half a year earlier, at Ignite 2025, Agent 365 was first presented. Last week, the Foundry team published the first technical documentation to answer “what is an autopilot?” and other related questions.

This is a great moment to do a reality check on the autopilots story. Just like the Copilot story and capabilities have been a long journey with grand promises and practical struggles to meet those expectations, I expect the road to autopilots to be less clear than the marketing message paints it to be. But let’s start from there and then go deeper into what we can observe in the actual Microsoft products.

In July, the CMO for AI at Work, Jared Spataro, posted a newsletter issue titled “Autopilots have landed. Here's what you need to know”. The nice and seemingly clear part of it was this graphic of the road to autopilots:

The chat vs. reasoning leap is pretty clear for anyone who’s seen the change in how ChatGPT, Claude et al. answer your questions. Instead of immediately starting to predict the next token and produce a visible output, they “think” internally for a bit, checking on their own work, and then show a much more refined answer. In exchange for spending more tokens, we’ve gotten more useful LLM-based services. There’s of course more to it behind the scenes, but let’s keep things simple as we’re in the marketing stage now.

Agent harness is a term that is now becoming something even normal end users of Microsoft technology have to figure out, thanks to Copilot Studio making it a visible and highly impactful choice in the new UI. I’m not sure if the Choose a harness page on MS Learn will make anyone say “Great, now I get it!”, but it is what it is. The technical implementation details keep bubbling up from beneath the UI layer, causing all kinds of branding and billing complexity. Whether you slap a Cowork or Codex or [whatever] Agent label on it, it’s basically about giving AI a computer with tools, memory, and a “manager” agent that ensures the given task gets completed somehow.

Finally, we get to the autopilots. Here’s what Jared’s post says: “Until now, every agent required a human to start the work. Autopilots don’t. Give an autopilot a goal and clear boundaries, and it will pursue that goal on its own, including approaches you never thought to ask for.” Okay, I guess Jared hasn’t heard of the flow triggers that Copilot Studio agents have had for quite some time, allowing them to get started by any trigger Power Automate supports. But let’s not get stuck on the details.

Based on the examples given in the picture, obviously Microsoft’s vision of an autopilot is modeled around OpenClaw—the thing Satya Nadella initially compared to a virus but later gave a new team the green light to pursue as an M365-enabled agentic product, eventually branded Microsoft Scout.

The promise of how “each step moves AI deeper into the fabric of work” paints this as an evolutionary path that creates ever more sophisticated AI services. Such agents will naturally require more sophisticated development, administration and governance tooling as well.

One week ago, a content drop on the MS Learn docs for Foundry gave us more interesting material to analyze. Specifically, the “What is an autopilot in Microsoft Foundry” page.

Let’s look at the definition as stated by the Foundry team. More specifically, the part “What doesn't define an autopilot”. Sure, this is obviously LLM-written text that has the usual Claudisms in its structure, but someone has signed off on it, so that makes it official.