The Windstorm

A weekly dispatch from the eye of the storm

Week of August 24, 2026

Welcome to the Eye of the Windstorm.

The most important AI number this week is a failure rate. Gartner is holding to its call that more than 40% of agentic AI projects will be canceled by the end of 2027, and IDC's count still has only around 12% of agent pilots reaching production at scale. Read that again. Somewhere between eight and nine of every ten pilots that work in the demo never make it into the business. The technology performed. The organization couldn't catch it. That gap is the whole story, and it's a human one.

1. The graveyard between pilot and production

Here's what kills these projects, and notice it isn't the model. Up to 54% of agent pilots stall three to nine months after a “successful” launch. A pilot that costs $200k to build can cost $1 to $2 million a year to run reliably at scale, and most budgets never modeled that. Data preparation eats 60 to 75% of the effort. None of those are intelligence problems. They're wiring problems, ownership problems, plumbing-and-people problems. The tools are the easy part. They always were.

2. The plumbing quietly standardizes

On August 20, Google's A2A protocol formally joined the Linux Foundation's Agentic AI Foundation, landing under the same neutral roof as Anthropic's Model Context Protocol. The foundation now counts more than 250 members: AWS, Anthropic, Bloomberg, Cloudflare, Google, Microsoft, OpenAI. Translation for the rest of us: how agents talk to your tools and to each other is becoming a standard, not a lock-in. Less integration friction, fewer bets you can't unwind. Boring news, and boring is exactly what infrastructure is supposed to be.

3. The models keep getting cheaper and faster

Google shipped Gemini 3.7 Flash on August 13; Z.AI, Alibaba, and others pushed out fresh models within days of each other. The race stopped being about the smartest model and became a speed race, a price war, and a distribution war all at once. Good for you: the raw capability sitting on your desk keeps improving while the bill keeps dropping. It also means capability is no longer your constraint. Which brings us to the eye.

The Eye

We've been here before. When electricity arrived, buying the current was never the hard part. Factories that simply bolted a motor where the steam engine used to sit got almost nothing. The ones that rewired the whole building around the new power, and taught every worker where the switches were, took roughly forty years to show up in the productivity numbers. The gap was never the electricity. It was the wiring and the teaching. This decade, forty years becomes four, but the shape is identical, and this week's failure rate is that shape made visible.

So stop asking whether the model is good enough. It is. You are a child monarch seated at a table of advisors who each know a thousand times what you do, and your only real power is the question. The great winnowing skill of this decade is knowing which advisor to ask, what to ask, and which of the answers to keep. Nobody is a virtuoso yet, which is the best news in this whole letter: the people who win the next few years aren't the ones with the best tools. They're the ones who learned to conduct them first. That's a human skill. You can start practicing it today.

Do this with your agent this week

Take one pilot or idea you're excited about and ask your agent a single question before you build anything: “If this works in the demo but dies before production, walk me through exactly how it dies—the cost, the data, the person who has to own it, the switch nobody flips.” Then go fix the answer, not the model. That's the winnowing skill in one move.

From the eye of the storm,
Grant