A project notebook for guiding MRI AI agents

AI agents can write and run code, but scientific work asks them to do more than produce an output without errors. In MRI, a reconstruction may run while misreading which data were acquired; a sequence file may need checks against timing, hardware limits and simulation assumptions. MRFoundry asks whether a structured collection of MRI knowledge and tools can help agents navigate such details, and reports how often agents met defined completion checks in a benchmark.

A toolbox, not a new MRI model

MRFoundry is an open-source package of “skills”: instructions that connect MRI concepts and references to software guidance, actions and output checks. The collection covers areas such as sequence design, reconstruction, hardware and diffusion MRI. It is intended for use with existing coding-agent environments; scientific applications and their dependencies must be installed separately.

The package also describes a project-local notebook for recording decisions, evidence, software versions, commands and failed attempts. Researchers can keep lessons local, adapt them, or propose supported changes to shared guidance for review. A method that fits one dataset or software version may not transfer, so the guidance emphasizes documenting scope and rechecking lessons in new settings.

The proposed workflow is a loop: a researcher sets the question and constraints; an agent consults guidance, uses tools and checks outputs; then the researcher reviews the evidence and decides what to do next. The authors describe recommended relationships, not an enforcement system that guarantees each step occurs.

Examples and paired benchmark

Two researcher-supervised examples show how the approach was used. In one, an agent built and exported a radial gradient-echo sequence, then simulated it with established software after researcher feedback. A later audit checked timing and trajectory and compared simulation results at two integration-step settings. These checks support claims about the saved file and computation, not scanner readiness, safety or complete physical-model accuracy.

In the other example, an agent inspected acquired brain-MRI data and distinguished measurements actually acquired from the larger stored grid before comparing reconstruction methods under shared settings. The revised L1 method reduced one image-discrepancy measure relative to initial SENSE, but its sampled-data residual increased. A zero-filled image remained closer to the noisy reference, and substantial data residuals remained. Closeness to that reference or numerical stability does not establish anatomical accuracy.

The benchmark compared paired agent attempts with and without the MRFoundry package. The intervention bundled skills, helper scripts and a discovery instruction; the study does not isolate the effect of guidance alone. Within pairs, the requested model label, task instance, inputs and host tools were matched, and runs used separate sandboxes. Task-specific time limits applied. The Core set covered 24 task types across ten models, with one instance per type and model, for 240 pairs. Hard comprised 134 pairs with unequal task coverage across model groups. A task counted as complete only when every frozen automated check passed; no language-model judge assigned completion.

Core estimate and Hard uncertainty

On Core, 188 of 240 attempts with MRFoundry passed every check, compared with 162 without it: 78.3% versus 67.5%, a difference of 10.8 percentage points. The reported 95% interval was 0.8 to 21.2 points. On Hard, the totals were 60 of 134 with the package and 49 without it: 44.8% versus 36.6%, a difference of 8.2 points. Its 95% interval, −3.9 to 19.7 points, includes no difference. Thus the Hard estimate is less certain; it does not establish an improvement.

These are observed differences under the tested conditions, not proof that the package will improve performance generally. Core results also varied by domain: diffusion completions fell from 17 of 20 without MRFoundry to 15 of 20 with it. The measure is success on benchmark checks, not research time saved, quality of scientific reasoning or correctness and usefulness in practice.

What remains unknown

Core used one run per task-model combination, so repeatability was not measured. Tasks were authored within an author-selected scope, independent MRI-expert review was incomplete, and one Hard image subcheck accepted a zero-filled reconstruction. Model labels do not verify exact provider-side model snapshots. Memory and disk policies changed during evaluation, and historical resource exposure was not fully established or uniform. Those limitations complicate interpretation of the comparison.

The examples were supervised, and the package-level comparison cannot identify which component accounted for the observed differences. Later reuse of notebook lessons and performance gains from community contributions were not evaluated. MRFoundry is best understood as a way to make MRI workflow knowledge easier to inspect and reuse—not as an autonomous researcher or a guarantee of valid results. This is an arXiv preprint; its findings should be read as preliminary and interpreted in light of the study’s scope and limitations.