Everyone’s shipping agents. The real job is the control plane.

The agent conversation has moved.

Six months ago it was still mostly demos and “look what Copilot can draft.” Now Microsoft is selling a control plane for agents — Agent 365 — so IT can observe, secure, and govern every agent across the organisation in one place.

That should tell you where we actually are.

When the platform vendor ships governance tooling as the headline, it means agents have left the novelty phase. They’re proliferating. They’re touching data. They’re creating cost. And without a way to see them, you’re not running an AI programme — you’re collecting risk with a friendly UI.

I spend my week inside Microsoft 365, Copilot, and the messy middle where clients try to turn licences into outcomes. What I’m watching closely isn’t “can the model do the task?” Anymore, that’s table stakes. What I’m watching is:

• Who owns the agent after the pilot applause?
• What data can it touch?
• What’s the real cost once usage isn’t a flat seat licence?Analysts are already forecasting a wave of agentic projects getting cancelled — not because the models failed, but because cost, value, and risk controls didn’t keep up. I’ve seen the same pattern in quieter form for years with Power Platform apps, SharePoint sprawl, and “transformation” that never quite landed. The tools change. The failure mode doesn’t.

So here’s my working rule as a CEO using this stuff myself:

Don’t measure how many agents you launched. Measure how many you’d defend in a board meeting — with the invoice open.If you’re a leader being sold agents right now, the useful questions aren’t feature questions. They’re operating questions:

1. What’s our registry — do we even know what we’ve deployed? 2. What’s our kill criteria before we start? 3. Who signs off spend when the meter starts moving? 4. What happens to the work when the agent is wrong on a Tuesday afternoon?

I’m going to keep writing that down here — what’s moving in the market, and what it feels like to run a real company through it.

Next up: a proper Week 1 ledger entry from inside the business. Setup, cost, hours back, verdict. No borrowed opinions.

The stack page on this site exists for that reason. Kept vs killed. Price on the page. No theatre.

• And do we have the courage to switch it off when it doesn’t earn its keep?

That last one matters more than people admit.