pcba-bench

pcba-bench is an agentic PCBA benchmark that asks how many manufacturable printed-circuit-board assemblies an AI agent can design from written briefs, unattended, in one hour. Ten boards range from a single LED to a BLDC motor controller, an SDRAM tester and a USB-C PD supply. Each is a short customer brief plus numbered requirements, and a run is one agent session with all ten briefs, a declared tool set and 60 minutes in total. The agent decides which boards to attempt and in what order.

For each board attempted, the agent must deliver a KiCad 10 schematic and PCB, Gerbers and drill files, a BOM with orderable JLCPCB and LCSC parts and a CPL, and a DECISIONS.md. Each attempted board is scored 0 to 100 on a six-part rubric by three judge models from three different vendors (Claude Opus 5.5, GPT-6 Astra and GLM-5.3), each in its own native harness and working blind from a deterministic evidence pack (KiCad ERC and DRC, copper-versus-schematic comparison, BOM resolution, renders); the per-criterion median is the official board score. Unattempted boards score 0, the suite score is the mean over all ten boards, and a separate Manufacturable count records how many boards would pass fab with no further work.