Remember OpenClaw? The circle has closed

Remember I wrote about OpenClaw earlier this year? What it was and what you could do with it?

Everyone saw somebody at a talk-show table speaking to his Clawdbot over WhatsApp. Early this year it went through the tech world like wildfire. For the first time a wide audience saw an agent that does not only answer questions, but actually does things for you and drives your PC.

Well, today that circle has closed.

From hobby project to enterprise

Microsoft announced an agent at Build built on exactly that same technology. Only in a suit.

It comes with a new category, which they call Autopilots: always-on agents that keep running in the background, with an identity of their own, taking action on your behalf. Including while you are busy with something else.

The nice part is that you give it a name and train it as you go. You give it feedback, it builds skills and memories, and gets better at how you work. The underlying memory learns what you consider important and what needs to happen next.

In practice: scheduling meetings across time zones, flagging important ones, preparing materials, blocking time for deadlines. And even spotting risks, like decisions that keep drifting.

And it does not run alongside your work. It sits inside the places you already are, and it works on your desktop and in your browser.

Why this is more than an announcement

When I wrote in June about agents working under your session, I named three ways an agent gets access: Borrow, Grant, Take.

Borrow is the agent that uses your session and has nothing of its own. That was OpenClaw on your own laptop, and it is still where most organisations are.

Grant is the agent that gets an identity of its own, with its own permissions and its own line in the log.

This is Grant. And that is precisely why it is more than a product announcement.

Every agent runs under its own managed identity. No anonymous, shared service accounts. Which means everything the agent does traces back to a known actor. Credentials are scoped to the task at hand, get redacted from logs, and sensitive actions can require manual approval before they proceed. Existing data policy (labels, data loss prevention) is simply enforced. The agent does not route around those controls; it works inside them.

That is the good part. The wild, untamed thing was fun at home, but you do not want an agent like that wandering through your corporate inbox unsupervised. Serious thought has gone into this.

And then what it does not solve

Two things, and they come in the same order as in my own framework.

Borrow does not go away. This sits next to it. On your employees’ laptops there are still extensions and assistants working inside their sessions. That does not diminish because a tidy variant now exists. When something becomes available in a proper form, the messy form does not usually retire.

And Grant brings its own bill. Hundreds of identities nobody cleans up, holding permissions nobody can still explain. We have been through this twice with service accounts. I see no reason it goes differently this time, unless you fix an owner and an end date per agent from day one. That is a boring administrative decision, and it is the only one that helps.

For now it is available through a preview programme, to a select set of customers, with licensing and configuration requirements around it. So this is not Tuesday yet. But it is the direction, and it arrived faster than I thought in June.

What I take from it

That a framework is only worth something once reality catches up with it. In June I wrote that Grant was the stage the market was moving towards.

Six weeks later it was announced.

That does not mean I called it right. It means the third stage is probably closer than I wrote down too: Take, the agent reaching access nobody granted it. And that is the part no vendor has a switch for yet.