The double diamond was built for human problems. It turns out AI agents have the same ones.

When we design AI agents, the temptation is to start with the capability. The model can do X — so we design for X. That’s the equivalent of skipping the first diamond entirely.

But the first diamond asks a harder question: what is the human actually trying to accomplish? Not what can the agent do — what does the person need to remain in control, stay oriented, and trust what’s happening?

An agent that surfaces the right answer at the wrong moment, in the wrong form, with no explanation of its reasoning — that’s a capability problem dressed up as a design problem. We built the solution before we understood the experience.

The double diamond reframes AI agent design around the human, not the model:

Discover — where does the human lose the thread in an AI-assisted workflow?
Define — what does trust actually look like at that moment?
Develop — how does the agent communicate its reasoning without overwhelming?
Deliver — how does the human stay in control when the agent is wrong?

The framework didn’t change. The stakes just got higher.

Designing for AI agents isn’t about making the agent smarter. It’s about making the human more capable alongside it.