As AI starts designing its own screens and flows, can we make that invisible decision-making visible again?

This explores the claim that AI moves design work into a hidden orchestration layer, and whether that layer can be made explicit. One caveat first: none of the retrieved notes covers Nielsen's orchestration-surface framing directly, so this answer can't tell you what his formulation says. What the corpus does have is several groups of researchers working on the same problem from different directions: when an AI decides what to build or do, where does the design judgment go, and how do you get it back into view?

The clearest version of the idea comes from HCI work on 'just-in-time objectives.' In that architecture, the system generates and evaluates its own goals at every interaction, and those steps are shown rather than buried. The argument is that making these choices visible is exactly what keeps human design expertise relevant as AI improves Does visible design structure keep HCI expertise necessary as AI advances?. The opposite case shows what happens when the layer stays hidden. Moran argues that generative UI is valuable because the AI acts as the designer for people who can't describe the interface they want Can generative UI serve users who cannot specify interfaces?. But a benchmark of generative UI tools found that roughly a quarter of stated design rationales never made it into the output, and the tools recognized only about half of the UX principles in their prompts Do generative UI tools actually implement their stated design rationales?. When design happens inside the model, it can quietly fail, and the result still looks finished. The benchmark's authors call this 'design theater.'

The engineering side reaches the same conclusion by a different route. Dr. Claw leaves the coding agent itself unchanged and wraps it in persistent state and skill libraries. That orchestration layer is what makes the work auditable and recoverable Can orchestration layers make coding agents more auditable?. GUI-agent research splits work the same way. Agent S separates high-level planning from pinpointing which on-screen element to act on Can structured interfaces help language models control GUIs better?. OmniParser converts a screenshot into a structured list of elements before the model chooses an action Why do vision-only GUI agents struggle with screen interpretation?. In both cases, an intermediate layer that someone deliberately designed outperforms asking the model to do everything at once. That intermediate layer is design work, just not the kind anyone draws in Figma.

There is a catch. Visibility alone isn't enough. Designers who could watch an agent's progress step by step did spot moments to step in, but when they edited the canvas, the agent misread or undid their changes, and they stopped trying Does visible agent progress help designers know when to intervene?. A visible orchestration layer still needs a way for people to edit it and have those edits stick. IBM's Carbon for AI shows how such a layer can be packaged as a reusable component: a styled AI label that opens an explanation popover. It also shows the gap, because the guideline reports no testing of whether the label changes user behavior Does consistent AI labeling actually reduce automation bias in practice?.

One thread you might not expect: multi-agent systems research formalized this kind of hidden structure more than twenty years ago. MOISE+ splits an agent organization into three separate specifications: roles and groups (structure), global plans (function), and permissions and obligations (deontic rules) Can separating structure and function improve multiagent organization design?. That three-part split is a ready-made starting point for anyone trying to formalize an orchestration layer today. Deloitte's argument that agentic AI fails when it is bolted onto old workflows, instead of redesigning the process end to end, is the business version of the same point Does agentic AI succeed through process redesign or workflow layering?. In every one of these sources, the orchestration is where the design actually happens.