I asked three AI agents a basic question about the same PIC GDS file:

“What are the components of gc_p0p618_n35_d0p55_011.gds?”

All three were “smart.” Only two were grounded. What changed the quality of the answer wasn’t model IQ. It was the agent’s relationship to the data: whether it could parse GDS, and whether it could query structured photonics metadata instead of guessing from names.

What each agent produced

1) OptixLog Knowledge Graph Agent (Opus 4.5)

This agent didn’t just list cells. It returned a component inventory as if it had a design database behind it, because it did.

Component inventory (5 categories):

• 2 cell instances
• 14 directional couplers (with specs)
• 2 GDS cells
• 2 geometries
• 34 waveguides

Quantitative details extracted:

• Waveguide dimensions: 342 nm width × 11,000 nm length
• Directional coupler gap: ~276 nm
• Geometry area: 307,125,000 nm²
• Layer assignments: Layer 1/0 (waveguides), Layer 2/0 (geometry)

Interpretation:

• Recognized this as a non-trivial photonic circuit structure
• Identified a grating coupler pattern