The Windstorm

A weekly dispatch from the eye of the storm

Week of September 21, 2026

Welcome to the Eye of the Windstorm.

This was the week the agents got names. On September 11, Salesforce shipped seven Agentforce agents with first names and job descriptions — Casey, Paige, Carter, Hunter, Marshall, Piper, and Fin — one apiece for service, sales, commerce, IT and HR, supply chain, and the rest. Not "an AI feature." A coworker with a nameplate. And here is what I keep circling back to from the calm center: naming the agent was the easy part. The model behind Casey is something you could not have bought at any price three years ago. Getting Casey to actually help your service team — that is the whole ballgame, and almost nobody has won it yet.

1. Salesforce gave its agents names and job titles

Casey answers your customers. Hunter chases leads. Marshall watches the supply chain. Giving a piece of software a name and a lane is a psychological move as much as a technical one — it tells your people "treat this like a hire, not a toggle." Good. Because you would never drop a new employee at a desk with no onboarding, no manager, and no clear job, and expect miracles. Yet that is exactly how most companies are deploying agents right now. The name is a promise. Whether it gets kept is up to you.

2. Accenture and Google hired 1,000 people just to flip the switch

The same week, Accenture and Google Cloud stood up the "Gemini Enterprise Business Group" — including a thousand forward-deployed engineers whose entire job is walking into companies and wiring the intelligence into the actual work. A thousand people. Not to build the model — the model exists. To wire the building. Remember when electricity took forty years to change the factory floor, and forty became four only once someone rerouted the whole line and taught every worker to flip the switch? Buying the electricity was never the hard part. This is the biggest tell of the week: the smart money is now betting on deployment, not models.

3. The great winnowing has begun

Now the cold water. Gartner projects that more than 40% of agentic AI projects will be canceled by 2027 — not for bad technology, but for unclear value and runaway cost. Roughly 97% of executives say they deployed an agent this year, yet only about 23% have scaled one into even a single function. And MIT's number still stings: 95% of enterprise pilots show no measurable return. But look at the other tail — the companies that get it right are earning $3.70 for every dollar spent. Same tools, opposite outcomes. The difference is never the model. It is whether a human asked the right question before turning it loose.

The Eye

Here is the quiet truth under all the noise: we have handed every company a table of advisors who each know a thousand times what any one of us does — and made every one of us a child monarch who has never learned to hold court. The agents got names this week. The people who direct them did not get any better at directing. That is the gap. Not compute, not context windows, not which lab shipped what on the first of the month. The bottleneck moved from the machine to the person sitting across from it, and it is going to stay there for a while.

So the skill of this decade is not prompting, and it is not coding. It is winnowing — knowing which question to ask, which answer to trust, and which of a thousand confident suggestions to throw out. It is the Socratic move: the monarch who rules well is not the one with the most answers but the one who asks the sharpest question and can smell a bad reply. Nobody is a virtuoso at this yet. Not the labs, not the consultants, not me. Which means the field is wide open — and the calmest person in the room, the one still asking questions while everyone else is shouting, is the one who wins.

Do this with your agent this week

Pick one task you keep handing your agent with the answer already baked in. This time, give it only the question and your real constraints — the budget, the deadline, the thing you are secretly worried about — and then ask it to interview you before it does anything: "What do you need to know from me to do this well?" Answer its questions honestly. You will learn more about your own thinking in ten minutes than in a week of typing instructions. That is conducting, not commanding.

From the eye of the storm,
Grant