Harvard physicist Matthew D. Schwartz has released BootLoops, an open-source toolkit that lets AI agents assemble tools for scientific calculations. In an Oct. 1 paper, he describes a model-independent package of programs that perform calculations and provide ways to test their outputs.
BootLoops is separate from the AI model. The paper says its contribution is to pool methods from physics, mathematics and computer science as working programs an agent can operate and extend; some programs adapt public code, while others were newly implemented. Project documentation says the toolkit can work with different AI agents.
In one Feynman-integral method, mathematical constraints narrow the possible form of an answer, and high-precision numerical samples determine its coefficients. The proposed formula is then compared with calculations at points that were excluded from every step of its derivation. In most cases, the reference values come from a separate computational route called auxiliary-mass flow.
The paper accepts a formula when the relative difference is below 10−30 at every held-out point; many results are checked to 60 or 100 digits. That is strong numerical evidence, not a symbolic proof.
The paper's reported examples focus on Feynman integrals and the Watson integral; its discussion of other fields does not amount to a broad validation study. Whether the toolkit generalizes will require independent tests on additional problems with clear acceptance criteria.
Schwartz says domain experts still choose meaningful questions and judge the significance of results. Anthropic identifies his essay as a guest contribution and says BootLoops is not an Anthropic project; Schwartz owns and maintains it.
