A peer-reviewed study describes a working architecture in which Claude Code autonomously operates optical pooled screening, a long-running biological image analysis method, across full experimental runs. The system follows a fixed sequence: a human approves a plan, the agent executes it, a monitor checks progress, and failures route back upstream for a human to resolve. The study reports measurable quality improvements over human analysts on the same task.
The architecture makes the human's role explicit at two points: before the run starts and when something goes wrong. Between those points the agent works without interruption, tracking its own progress and recovering from recoverable errors before escalating. That structure, approve-then-run rather than step-by-step supervision, is the central interaction design choice the study documents.
For product teams, the study is notable because it is empirically validated on a real scientific workload, not a benchmark. The quality improvements over human analysts give the approval-and-escalation structure a concrete outcome to point to. The design tension it exposes is how much autonomy to grant before the first human checkpoint, and how to define the boundary between errors the agent should absorb and errors that require a human decision.