The Windstorm

A weekly dispatch from the eye of the storm

Week of August 31, 2026

Welcome to the Eye of the Windstorm.

Ninety-seven percent of companies say they deployed an AI agent this year. Fifteen percent got it to actually scale. That gap is the whole story, and it is not a technology gap. New numbers this week put a finer point on it: of the custom AI tools enterprises build, roughly sixty percent get evaluated, twenty percent get piloted, and one in twenty ever ships to production. The tools work. The wiring between the tool and the human is where everyone is stalling out. Same as always: the tool is the easy part.

1. The plumbing I flagged last week just became official

The handover I mentioned last week is now done: on August 20 Google gave its Agent2Agent protocol to the Agentic AI Foundation, where it now sits beside Anthropic's Model Context Protocol under one neutral roof. That foundation went from 49 members to more than 250 in under a year — AWS, Anthropic, Google, Microsoft, OpenAI, Bloomberg, Cloudflare all signed. Translation: the industry just agreed on how agents talk to tools and to each other. The plumbing is done arguing. Which means the frontier moved — again — from "can we build it" to "can we use it."

2. Your agent can now read the web without leaking your data

The day after, August 21, AWS made Web Search on Bedrock AgentCore generally available — a managed tool that lets your agent fetch live, cited answers from the open web without a single byte leaving your own cloud account. A year ago that was a research project. This week it's a checkbox. Notice the pattern: the capability keeps getting handed to you for free, and the bottleneck keeps being the person deciding what to ask it.

3. The firehose is not slowing down

August brought two dozen new models from eighteen providers — GLM-5.3 Flash landed on the 26th, and it will be old news by the time you read the next one. The launch cadence has roughly quadrupled since 2023. Here's the uncomfortable part for leaders: nobody can keep up, and keeping up was never the assignment. Choosing well is. Knowing which of these to ignore is now a competitive skill, and almost no one has it yet.

The Eye

We keep telling the story of AI as a story about the machines. It is really a story about a child monarch. You now sit at a table of advisors, each of whom knows a thousand times what you do about their little domain — the legal one, the market one, the code one. You will never out-know any of them, and you don't need to. Your job is not to be the smartest voice at the table. Your job is to ask the question that makes the whole table useful, and then to decide. That is the entire skill. The advisors got infinitely better this week; the monarch did not.

Think of electricity. The science was settled for forty years before most factories saw real gains — because the hard part was never buying the power. It was rewiring the building and teaching every worker to flip the switch. We are in the wiring years right now. The company that wins the next decade will not be the one with the best model. Every model this week is astonishing and every one is a commodity. The winner will be the one that taught ordinary people to ask sharper questions and to winnow — to throw ninety-five percent of the output away and keep the five that matter. Nobody is a virtuoso at that yet. That is the good news. It means the seat is still open, and it is yours to take.

Do this with your agent this week

Pick one decision you're currently stuck on. Don't ask your agent to solve it. Instead type: "Ask me the five questions I should be able to answer before I make this call." Then sit with the questions it hands back. You'll learn more from what you can't answer than from anything it could have told you.

From the eye of the storm,
Grant