The EDA industry is racing toward AI agents for physical design: place and route, signoff, everything in between. The more we build, the clearer it becomes that the quality of the model was never the issue.
An agent that acts on its own has to reason across domains and get feedback quickly. Three things make that genuinely hard today.
The flows are multi-vendor by nature
Placement, congestion, timing, power integrity, thermal, stress: each tends to come from a different tool, often built for much older process generations. An agent working across them sees fragments, never the whole.
Nothing here happens in isolation
A small ECO can ripple into congestion, timing, IR drop, and thermal effects you didn't see coming. When a single tool run takes days, you don't get many chances to iterate. Make a fix without a cross-domain view and you can spend the next week chasing the side effects it caused.
The sheer volume
A modest block can hold millions of gates, hundreds of millions of shapes, millions of voltage-drop values, and gigabytes of timing reports. Pour that raw data into an LLM and you blow past the context window, burn tokens, and still leave the model with no idea where to look.
We keep coming back to one mental image. Running separate agents for timing, IR, and congestion, each reasoning on its own, is like putting three engineers in a room where one only knows timing, one only IR, one only congestion, and asking them to agree on a decision. You rarely get the best answer that way. The designers who make the great calls are the ones holding the whole picture at once, weighing the tradeoffs in their head simultaneously. An agent should be measured against that same standard.
So when people ask whether a better model will get us there, our honest answer is: not on its own. A model is only as good as what it can actually see, and right now it can't see much. That's the part we think is worth fixing first.