BCER Agent brings reliability to long-horizon MRI workflows with bounded recovery

Real MRI analysis requires long, interdependent pipelines on 3D/4D volumetric data, but reactive tool-calling agents often fail due to cascading breakdowns from faulty references or mismatched arguments. To address this, a team led by Ziyang Long (Cedars-Sinai) and including researchers from multiple institutions developed BCER (Brain-Cerebellum-Extremity-Reflector), a controller architecture for reliable long-horizon MRI workflow execution. BCER separates high-level planning from execution and introduces bounded local recovery, allowing the system to handle errors within a limited scope without restarting the entire pipeline. The approach was evaluated on a multi-organ MRI benchmark covering brain, prostate, and cardiac tasks with both short- and long-chain workflows.

Results show consistent improvements over reactive baselines, with the largest gains on long-chain workflows. BCER also enables auditability by maintaining explicit links between final outputs and intermediate artifacts and measurements. The code and benchmark have been released. The paper has been accepted to MICCAI 2026. This work represents a significant step toward deploying autonomous AI agents in clinical MRI analysis, where reliability and traceability are critical. The bounded local recovery mechanism is particularly important for medical settings where full re-execution is costly or time-prohibitive.