An AI system can now write, validate, and revise wet-lab protocols well enough to pass device-level execution checks nearly nine times out of ten.
Researchers introduced ProtoPilot, a self-evolving multi-agent system designed to translate biological protocols into runnable lab automation code. Tested against a benchmark of 294 tasks drawn from 98 gold-standard synthetic- and molecular-biology protocols, the system achieved an overall protocol-to-code gate pass rate of 89.5% and an Opentrons-specific pass rate of 88.24%. For context, OpenTrons-AI — the existing dedicated tool — cleared just 32.35% of the same Opentrons tasks. Wet-lab runs produced Sanger-confirmed DNA products and feedback-corrected assemblies, meaning the outputs were verified in physical experiments, not just simulated.
The gap between ProtoPilot and OpenTrons-AI is wide enough to matter: more than doubling the pass rate on real instrument code suggests the multi-agent architecture — which layers protocol generation, SOP expansion, SDK-compliant code synthesis, and runtime skill updates — handles device constraints in a way that single-purpose tools do not. A 90.2% expert-preference rate on head-to-head comparisons adds a qualitative signal on top of the quantitative gate metrics.
Wet-lab automation has long stalled at the "plausible but unrunnable" stage, where AI can describe a procedure but cannot reliably translate it into instrument commands. ProtoPilot does not solve biology; it narrows the gap between a written protocol and a robot that can actually execute it — which, if it holds outside benchmark conditions, is the part that has always been the bottleneck.