2026 is the year agents actually arrived.

Last year was supposed to be that year. Instead 2025 was the year the pieces quietly snapped into place. Models got smart enough to reason through real work. Coding agents proved it could work in the real world. The platforms followed. At this year's Google Cloud Next, Vertex AI was renamed the Gemini Enterprise Agent Platform: a clear signal that the conversation has moved from "AI models" to "agents in production". The other hyperscalers are telling the same story. The infrastructure is real, governable, enterprise-grade. It had to exist before any of this could graduate from pilot.

So what's on the table? It helps to be clear about what we mean, because right now "agent" gets stretched to cover almost anything, much like "AI."

Assistive agents:

The ones that work beside you, like a sharp colleague in constant back-and-forth. They give your people leverage on the work they're already doing.

Autonomous agents:

The ones you send off to do a job on their own, the way you'd hand a project to a capable intern and check in along the way. They give you velocity on work that used to demand someone's full attention.

Infrastructure agents:

The ones running quietly inside your operations, doing work that used to pass through a person or a tangle of brittle automations. They let you redesign entire processes, not just speed up the ones you have.

The first is the most familiar and the easiest to picture. The other two are where the deepest leverage lives, and where the biggest business outcomes will come from.

Technology isn’t the blocker.

The thing standing between you and a useful agent today isn't the technology. It's that you can't hand off your work until you can describe how you actually do it, and what "good" looks like when it's done. Most of us can't. Not because we aren't smart. Because we're good at our jobs.

Think about it. The judgment that makes your best people valuable is the stuff they do without thinking. That instinct is the asset. It's also invisible, built over years, living in people's heads, rarely written down. It's exactly what an agent needs from you. And the single hardest thing to put into words.

That's taste. It's your edge, but it's also the bottleneck.

Articulation is the challenge.

This is the part the platform vendors can't solve for you. They'll deliver powerful platforms and models. But "AI as your head of strategy" isn't a product you can buy. Codifying how your business actually works, what your people care about, what a great outcome looks like for you, is not a configuration setting. It's a craft. The good news: craft can be learned. Companies that start the work now will pull ahead of the ones still waiting for a product to do it for them.

Defining what "good" looks like is the crux of that craft. Back in 2024 I ran a small experiment to see if an AI model could look at a piece of creative and tell you whether it was on-brand. The setup was almost embarrassingly simple: a brand's guidelines, a few examples of work that hit the mark, and a question. It worked better than I expected. The real surprise wasn't the verdict. It was the explanation. The model could tell you why something missed and how to fix it, in plain language a non-designer could act on. The clearer I made the bar, the sharper its judgment became. An agent is only as good as your ability to show it what good looks like.

That experiment taught me something bigger than "AI can check your brand." The static document was always a workaround. Brand guidelines existed because we had no other way to scale on-brand judgment. So we wrote the rules down, picked some examples, and hoped people would internalize the rest. Once an agent can carry that judgment directly, looking at the actual work and explaining what's off and how to fix it, the document was never the destination. The capability was.

Don’t bolt agents onto old workflows.

Most companies miss this move. They take the workflow they already have and ask where to slot an agent in, bolting one onto whatever steps already exist. But many of those steps were workarounds for the same fundamental problem: there was no other way to apply good judgment at scale. Take that constraint away, and the shape of the work itself starts to change. The better question isn't where to put an agent. It's what the work should look like now that the constraint is gone. When autonomous intelligence is cheap and on tap, you design backward from the outcome you actually want, instead of paving over the path you're already standing on.

The judgment lives at every step, whatever those steps are. Knowing what "good" looks like is only half of that. The harder half is how you get there. An agent works in a loop, making dozens of small judgment calls along the way: what to check first, which thread to pull, when something feels off, when to stop. That sequence, your method in motion, is taste too. It sits deeper than the outcome, because you run it on instinct. Naming it means putting years of intuition into words.

This is why the three types of agents demand different things from you. Assistive agents need you to recognize "good" when you see it. Autonomous and infrastructure agents need both halves: that recognition plus your method in motion. That's the harder excavation. It's also where the biggest business outcomes are waiting