The Windstorm

A weekly dispatch from the eye of the storm

Week of October 12, 2026

Welcome to the Eye of the Windstorm.

Last Thursday, Google Cloud stood up at its Gemini at Work event and launched a single “universal agent for work.” One agent, reachable from the web, your phone, the command line, and — this is the part to sit with — inside Google Workspace, Microsoft 365, and Slack alike. You don’t tell it which buttons to press. You tell it what you want accomplished, and it works out the steps and hands you back the finished thing, choosing between Google’s own models and Anthropic’s Claude depending on the job. For three years the honest excuse in most companies was “the technology isn’t quite there yet.” This week, quietly, that excuse expired. The capability is now ambient — it’s in the room whether you asked for it or not. Which leaves exactly one unsolved variable in the whole equation: whether you can say, plainly and well, what you actually want done.

1. Google collapsed the toolbox into one coworker

Here’s what changed on October 8. For years, “adopting AI” meant stitching together a dozen apps, each with its own login and its own learning curve — and most of that friction was deployment, not intelligence. Google’s new Gemini agent erases that seam. It’s in private preview, headed to select Business and Enterprise plans, with finance and legal versions already in preview. Notice what Google is really selling: not a smarter model, but a thing you delegate to. The wiring is done. The only question it can’t answer for you is the one you have to bring: delegate what?

2. Microsoft built the fence the same week

One day earlier, on October 7, Microsoft made Execution Containers generally available — OS-level sandboxes that decide what an agent is and isn’t allowed to touch, enforced by the operating system itself so the agent can’t quietly grant itself more room. Developers and IT administrators set the policy: which files and network destinations an agent may use. The risks get named out loud — UI spoofing, cross-session data leakage. The timing is the lesson. The accelerator and the brakes shipped in the same 48 hours, because a capable agent with no fence isn’t power, it’s exposure. Deciding where the fence goes is a judgment, not a feature.

3. 79% are already running them. 2% can govern them.

On October 6, SailPoint published its Horizons of Identity Security survey of 340 senior leaders, and the gap it found is stark: 79% of enterprises now run AI agents in production, while only 2% use identity security tools built specifically to govern them. Read those two numbers next to each other and you can see the whole problem of the moment. The agents didn’t arrive too slowly. They arrived faster than our ability to supervise them — which is another way of saying the bottleneck was never the machine.

The Eye

Step back from the headlines and look at the shape of the week. In a single news cycle, two of the largest companies on earth shipped opposite things: Google shipped an agent that can do almost anything you ask across almost every app you own, and Microsoft shipped the cage to keep that same kind of agent from doing things you didn’t. The power and the restraint, published a day apart. That is this entire decade compressed into 48 hours — because both halves are now solved problems that vendors will sell you off the shelf. What no vendor can sell you is the thing sitting between them: the judgment to state, precisely, what you want done and how far the machine may go to do it.

And here is the quietly encouraging part. That judgment — the skill of conducting a near-omniscient tool through one well-formed request — is maybe three years old. Nobody is a virtuoso at it yet. There is no twenty-year veteran of asking a universal agent the right question, because universal agents are a week old. So the people who will run the next decade aren’t the ones with the most credentials or the biggest model budget. They’re the ones who get good, fast, at the oldest skill there is: asking a sharp question and knowing what a good answer looks like. The storm out there is all noise about the tools. In here, the work is simpler and harder — learn to ask.

Do this with your agent this week

Pick one recurring task you’d say you already “know how to do.” Instead of doing it, write the objective as a single clear paragraph — the kind a sharp stranger could execute with no further questions. Hand that paragraph to your agent and watch where it stumbles. Every place it gets confused is a place your thinking was fuzzy, not the machine’s. Tighten the paragraph, not the tool. That gap — between what you can do by hand and what you can clearly ask for — is the exact gap this decade pays for.

From the eye of the storm,
Grant