This agent reads a cancer patient's file, molecular report, and scans, then recommends a treatment plan — the job clinicians call a molecular tumor board. It was rebuilt faithfully, then each of three subsystems was swapped out one variant at a time. Pick a subsystem to see how far each design choice moves completeness: the share of a case's clinician-written criteria the answer covers.
Completeness on ferber20 (20 cases, 109 clinician criteria), averaged over 5 runs. Dashed line marks each subsystem's baseline. ★ marks the choice the decomposition recommends.