Ask an experienced physical designer how they actually spend the day and you'll usually hear: reading reports, staring at summary dashboards. It's not that reports are the best way to build a feel for a design. It's that the classic GUIs can't keep up. A large block can take hours just to open, and a true full-chip render is basically off the table with the current EDA design tools.
We've spent a lot of words on the need for a unified intelligence layer that can answer multi-domain, multi-physics questions fast, one that ties timing, IR, congestion, and thermal together instead of leaving the agent to stitch them itself. Even with that in place, showing the agent what's actually going on adds something a table can't: the spatial intuition a human picks up from a heatmap or a layout. The classic line still holds: A picture is worth a thousand words.
Multimodal AI is what finally makes that practical without the GUI tax. Render the same congestion map, thermal plot, or IR overlay a designer would eventually pull up if they had the patience, and let the agent look at it directly. Vision models have gotten genuinely good at this kind of thing: spotting a hotspot, a routing pinch, an asymmetry that shouldn't be there.
The thing standing in the way has always been the GUI itself. You need one that can render any data, at full fidelity, with no simplification, in roughly a second, no matter the scale. That means going from a single block to a full chip to a 3DIC stack without trading speed for detail, including real 3D viewing of layout and heatmaps that still resolves down to the transistor. Make visualization that fast and it stops being a bottleneck and starts being a real input for an agent, the way it always should have been for humans. A side benefit: a GUI that quick and that easy to use turns out to be a genuine pleasure for the people still in the loop, too.
At 3DIC the picture has to cover a stack of dies, with thermal and power coupling running between them. Slow, heavyweight tools already make that hard for humans to see. A layered, real-time view gives an agent something closer to what a 3DIC engineer works from, instead of a pile of per-die numbers with no spatial link between them.