Design Conductor

Hand a chip-design agent a 219-word requirements document, leave the room, and come back twelve hours later to a finished processor layout. That is what Verkor's Design Conductor did: starting from a short prompt, it wrote the logic, built its own tests, debugged the failures, squeezed the timing, and drove the result through to a tape-out-ready layout — the file a foundry would etch — with no human touching the work in between.

"Basically, we are trading off experience for compute." — David Chin, VP of Engineering, Verkor

Earlier AI tools automated single stations on the chip-design line: one optimizes the floorplan, another searches the parameter space, and engineers stitch the stages together. Design Conductor is the first to run the whole line itself, spec to layout, as one agent. The chip it produced, called VerCore, holds its timing at 1.5 GHz and scores about where a budget Intel laptop chip from 2011 landed.

The catch is in how it got there, and it is the most revealing part. The processor is a simple one — a plain in-order design with no caches, the kind a competent engineer would call routine — and the agent still burned tens of billions of words' worth of computation to reach it. When it hit a timing bug, it didn't diagnose the cause the way an engineer would; it thrashed, making broad changes and hunting blindly through dead ends until something stuck. It was also never actually manufactured, only checked in simulation.

So the machine that designed a chip overnight is, up close, trading experience for raw compute — and the trade gets worse fast, because the search space explodes as the chip grows more complex. The question this opens isn't whether an agent can design silicon. It's how far brute force climbs before judgment is the only thing left that scales.