The Windstorm

A weekly dispatch from the eye of the storm

Week of September 28, 2026

Welcome to the Eye of the Windstorm.

Last Thursday a company announced software whose only job is to find the AI agents your own organization is already running — and can't account for. Dataiku called it Agent Management, and the problem it names is “agent sprawl”: your people are spinning up agents faster than anyone can track them. Sit with that. The frontier is no longer “can the machine do it” but “do you even know what your machines are doing.” The capability was never your bottleneck. Knowing where you pointed it always was.

1. The tools got better and cheaper again — quietly

Last Tuesday, Anthropic shipped Claude Opus 5.5 and, about an hour later, OpenAI shipped GPT-6 Sol and Luna, its cheaper tiers. Opus 5.5 posts frontier coding results at roughly 40% less cost than the last generation on typical work, with a million-token context window. A year ago that would have owned the headlines for a week. This time it registered as weather. Raw capability has become so abundant it barely moves the needle. What's scarce isn't intelligence. It's the judgment to aim it.

2. The money stopped pretending to be small

In September, Gartner raised its 2026 forecast for worldwide AI spending to $2.67 trillion — up 49.5% in a single year. Read the number underneath it: for every dollar the world spends on generative AI models themselves, it spends more than fifty-two on infrastructure — servers, power, cooling, networking. We are pouring an ocean of concrete for a machine most companies have not learned to operate. Spending is not deploying. Deploying is not results.

3. The agents got loose

Back to Dataiku. Their tool, generally available next month, scans across AWS, Google, Microsoft, Salesforce, Snowflake and more to inventory every agent a company runs, then measures whether each one delivers real business value — not merely whether it runs without errors. That second part is the whole game. An agent that runs flawlessly and helps no one is worse than none: it costs money and it hides. This is the great winnowing, right on schedule. The skill of the decade is deciding which of your thousand willing helpers earns its seat — and letting the rest go.

The Eye

Here is what all three stories are really saying, read from the calm in the middle of it. The tools crossed the “good enough” line a while ago. The money crossed the “serious” line this month. And the only line still uncrossed — the one every headline keeps tripping over — is human. A company cannot find its own agents not because the software is bad, but because nobody decided who owns them, what they are for, or how you would know if they worked. That is not a technology failure. It is a leadership vacuum that technology walked into.

So the most valuable thing you can bring to this is not a bigger model or a bigger budget. It is a better question. When someone shows you an agent, do not ask “does it work?” — everything works now. Ask “what would it look like if this were actually helping, and how would we see that number move?” The person who can ask that — precisely, out loud, before the money is spent — is worth more this year than the person who can build the thing. In a world where the machines can answer almost anything, the one scarce and decisive skill left is knowing what to ask. That is the whole job now. That is conducting.

Do this with your agent this week

Pick one agent, assistant, or automation you already lean on. Before you touch it, write down the single number that should move if it were truly earning its keep — hours saved, tickets closed, deals advanced. Then ask it, out loud: “If you were doing this job perfectly, what would be different by Friday, and how would I measure it?” If you can't answer in one clean sentence, you haven't found a tool problem. You've found the work.

From the eye of the storm,
Grant