AI agents can now route circuit boards — badly, then better, then almost as well as the tools engineers already use.
Researchers released PCBWorld, an open-source benchmark environment that drops AI agents directly into the KiCad EDA engine to handle PCB routing — the job of connecting a board's electrical nets with copper traces while staying inside strict design rules. The environment lets both reinforcement learning policies and large language models interact with KiCad's native operations and receive real-time feedback from its Design Rule Check system, the same tool human engineers use to catch layout violations. A companion dataset, PCBWorld-Bench, includes 679 real open-source boards plus synthetic instances, scored across eight engine-verified metrics. In experiments, agents working interactively inside the engine beat grid-action RL policies and open-loop LLM baselines — and an RL policy trained entirely on synthetic boards transferred zero-shot to real boards, getting close to the performance of rule-based routers.
PCB routing is one of those problems that sounds mechanical but resists automation: the search space explodes with board complexity, and a single constraint violation can scrap a design. The fact that a synthetic-trained RL agent can close most of the gap to rule-based routers — without seeing a real board during training — suggests the engine-in-the-loop approach is doing real work, not just benchmarking against a soft baseline.
That said, "approaching rule-based routers" is not "matching them," and production boards are rarely the tidy open-source kind. The benchmark is a foundation, not a finish line.