Anthropic's AI ran a biology lab's image analysis start to finish

A peer-reviewed study describes a working system where Anthropic's AI coding tool runs optical pooled screening across full experimental runs. Optical pooled screening is a biology method that processes large sets of microscope images. A human approves a plan at the start. The AI then runs the full job, tracks its own progress, and recovers from small errors on its own. Failures the AI cannot recover from route back to a human to resolve. The study reports measurable quality improvements over human analysts on the same task. The study tested this on a real scientific workload, not a constructed test. The quality improvements give the approve-then-run structure a concrete outcome to point to. The central design question is how much work the AI should do between the human's first approval and the next check-in. A second question follows: which errors the AI handles versus escalates to a human.