David Rostcheck got a coding agent from a PyTorch model to running circuitry on an Amazon Web Services (AWS) F2 field-programmable gate array (FPGA), a chip whose logic can be rewired after manufacturing. The useful result isn't speed. His partial TinyBERT implementation is slower than a CPU or GPU because it streams weights through host memory. The useful result is a four-level test ladder that catches a bad number before it becomes an expensive cloud build.
The same operation runs first in full-precision PyTorch, then in bit-accurate quantized Python, then in a C simulation of the generated circuit, and finally on the physical FPGA. Each boundary narrows the search. That separation exposed a generated exponential function that returned zero only after the design reached real hardware, plus a divider that couldn't meet a 250 megahertz clock. The paper also records an honest limit: only the embedding stage and one of TinyBERT's four encoder layers reached the board.
This is a better pattern for AI-assisted hardware than asking whether an agent can produce code that builds. A build can preserve a numerical error through hours of synthesis. Our bet is that the useful agent benchmarks in 2027 will report which validation stage caught each defect and how many defects escaped to the board, not just whether the final design compiled.