Every agent console I've seen this year is a wizard in a trench coat

Step through the intent. Review the plan. Run the agent. Approve the output.

Four steps, linear, with the agent doing the middle two and the human bookending. It's the exact interaction model we were taught to avoid in 2010.

There's a reason it keeps reappearing. Agents are still mostly linear under the hood, and the wizard is the pattern that makes a linear process feel reviewable. It's a safe default.

The problem is that it bakes in assumptions that don't hold.

That the user can articulate the goal upfront. That the plan is the right review surface. That work proceeds in discrete, approvable stages. That the human's job is to gate, not to shape.

Real agentic work doesn't look like that. It looks like iterative refinement, parallel exploration, mid-flight correction, and shared authorship of the plan itself.

The teams that break out of the wizard first will define what agentic UX actually is. The rest will be stuck shipping Clippy with a language model.