The Windstorm

A weekly dispatch from the eye of the storm

Week of July 13, 2026

Welcome to the Eye of the Windstorm.

This week Cisco handed an AI agent to every single one of its 90,000 employees, and in doing so handed the rest of us the clearest lesson of the year. Buying 90,000 agents is the easy part. You sign a contract and it's done by lunch. The hard part is the 90,000 people who now have to learn what to do with the thing sitting on their desk. We keep counting the agents. Almost nobody is counting the humans who have to learn to conduct them. That gap — not the model, not the money — is the whole game this decade.

1. Cisco gives all 90,000 employees a personal agent

Starting the end of July, every Cisco worker gets a personalized agent that routes each request to whichever underlying model fits the job, much of it running on-premises for cost and data control. The tool is astonishing and almost beside the point. What's telling is what Cisco is spending its worry on: company-wide upskilling, internal knowledge-sharing, and a push to get people to actually experiment. They already have proof it pays — the CFO says 80 to 90 percent of the first draft of their public financial filings is now AI-assisted. And in the same window, Cisco is cutting close to 4,000 jobs. Read those two facts together. The agent didn't create value or destroy it. The people wielding it did.

2. The tools got easier again — ChatGPT Work and Claude Cowork

On July 9 OpenAI launched ChatGPT Work, pairing chat with its coding engine so non-technical people can build spreadsheets, decks, and small web apps by asking. Two days earlier Anthropic pushed Claude Cowork to web and mobile — long-running jobs that keep going while your laptop is closed, turning a folder of contracts into a renewal tracker overnight. Add GPT-5.6 arriving as three tuned models. Notice the direction: agents that complete work, not agents that answer questions. Which means the scarce skill is no longer typing the prompt. It's knowing which of the ten things your agent hands back is the one worth keeping. That winnowing is the skill of the decade, and nobody — nobody — is a virtuoso at it yet.

3. Everyone's in production. Almost nobody's capturing the value.

The numbers this month tell one story twice. Gartner projects 40 percent of enterprise applications will have embedded agents by year-end, up from under 5 percent in 2025. Roughly 72 percent of agent projects are now in production. And yet a TEKsystems study lands the punchline: high adoption, low value capture, with 95 percent of IT leaders reporting integration problems. The agents moved to production and the organization got left behind. This is not a technology failure. It's an org-chart failure, a workflow failure, a training failure — a human failure wearing a technology costume.

The Eye

Here is the quiet thing under all the noise. The frontier labs themselves have stopped pretending the model is the bottleneck. OpenAI stood up a whole deployment company, over four billion dollars, and bought a consulting firm just to embed engineers inside other companies and rewire their workflows by hand. Sit with that. The people who build the smartest tools on earth are now spending their capital on the un-glamorous, deeply human work of getting organizations to actually use them. They've conceded the point: the intelligence is a commodity; the conducting is not.

Think of yourself as a child monarch seated at a table of advisors, each one a thousand times more capable than you in their domain. You will never out-know them. That was never the job. Your job is to know which advisor to turn to, what to ask, and — hardest of all — which answer to trust. A hundred years ago we spent forty years learning to wire the electricity we'd already bought; the switch was cheap, the wiring and the teaching were everything. We're compressing those forty years into four. The tool is done charging. Whether it lights your building depends entirely on the person standing at the switch.

Do this with your agent this week

Pick one recurring task you own end to end — a weekly report, a batch of emails, a status update — and hand it to your agent in full. Then do the part almost everyone skips: grade the result like a manager, out loud. What's wrong? What did it assume? Ask it the Socratic follow-up: "What would you need from me to do this better next time?" The prompt is the easy part. The question is the superpower.

From the eye of the storm,
Grant