A new multi-agent AI system is trying to automate one of chip design's most stubborn bottlenecks: analog and RF circuits.
Researchers built RFChipAgent, a flow of large language model agents that collaborate to design analog and RF circuits under human supervision. The system pulls engineering knowledge from existing documents using a retrieval system, picks circuit topologies with a dedicated agent, and builds schematics and testbenches automatically. A hybrid optimization engine then tunes component values, running every candidate through a circuit simulator and logging trusted results to guide later attempts. The team tested it on 60 GHz wideband low-noise amplifiers built on GlobalFoundries' 22FDSOI process, the kind of front-end circuit that Wi-Fi 7 and 6G radios depend on.
Analog and RF design has resisted automation longer than digital logic because it runs on messy physical tradeoffs - noise, power, linearity - that don't reduce neatly to code. This flow leans on a simulator to check every agent's guess before trusting it, a more disciplined approach than letting a model sketch a schematic and hope it works. That matters because Wi-Fi 7 and 6G standards are pushing more of this analog work onto designers, not less.
The paper's own numbers - a "substantial" cut in design effort on one amplifier family in one process node - are a promising start, not proof the RF engineer's job is done.