The Windstorm

A weekly dispatch from the eye of the storm

Week of October 26, 2026

Welcome to the Eye of the Windstorm.

On Monday, Microsoft did something odd for a company that sells AI: it told you not to trust it. In a company blog post, security chief Hayete Gallot wrote that “frontier models, regardless of their origin, cannot be trusted” — not because they’re evil or easily hacked, but because they’re unpredictable, trained to finish the task and not to honor your company’s judgment. Earlier this month, in an essay on X, Satya Nadella urged leaders to treat frontier models as “insider risks,” reminding them that “we simply can’t outsource responsibility for what intelligence does on our behalf.” The most powerful software vendor on earth just conceded that capability is no longer the question. The question is whether you can supervise it — and that one is entirely human.

1. Microsoft shipped a babysitter for your software

Alongside the warning came the tooling: Guardian Agents to watch the agents you run, a Defender update that now spots some sixty kinds of agents employees have quietly dragged into work (Claude Code, Codex and GitHub Copilot among them), permissions that phone a human when an agent reaches for something risky, and a seat in NVIDIA’s month-old agent-safety platform, which can stop agents at the hardware level — something like a kill switch. Strip the product names away and the message is one sentence: manage an agent the way you’d manage a brilliant, fast, slightly feral new hire. Give it the keys it needs and not one more. Watch what it does.

2. Oracle made building your own agents a no-code afternoon

Two days later, Oracle announced its AI Agent Studio for its industry applications — a no-code, low-code workbench for building and orchestrating whole teams of agents tuned to your business, with prebuilt industry agents to start from and U.S. availability planned within a year. Each agent inherits the access of the person running it and leaves an audit trail, so it can’t see what you can’t. Building the thing used to need engineers. Soon it needs you to know what “good” looks like in your shop. The bottleneck slid from can we build it to do we know what we want it to do.

3. A startup hit $1B betting everyone gets a teammate

Town raised a $108M round at a $1 billion valuation, co-led by Forerunner and Menlo, four months after launching “Townies” — an assistant for each employee that now hands tasks off to the others. Paying customers up fourfold since July. The money isn’t the story. The shape is: the org chart is quietly becoming an orchestration problem.

The Eye

Here is what the noise buried this week. The biggest names in the business all pivoted on the same beat — away from selling you more intelligence, toward selling you control of it. Guardrails, audit trails, permissions, a kill switch. When vendors start competing on how to supervise the machine instead of how smart it is, they’re telling you something they’d never put on a billboard: the model is finished. You are the variable now.

Which means the defining skill of this decade isn’t prompting, and it was never a trick. It’s the oldest skill in management — delegation with judgment. A captain can’t stand every watch himself, and doesn’t try; he decides what “done right” looks like, asks the one question that surfaces trouble early, and knows the moment to stop and check. That’s the whole job with agents, too. Not doing the work — knowing what good looks like, and asking better questions than the machine ever will.

Do this with your agent this week

Pick one task you’d normally just hand off. Before you let the agent run, write the three questions you’d ask a new hire you didn’t fully trust yet: What does “done” actually look like? What are you not allowed to touch? And what should make you stop and come ask me? Paste those answers into the prompt as the guardrails, and watch the job change under your hands — from doing the work to directing it.

From the eye of the storm,
Grant