How BootLoops uses AI agents to solve cross-disciplinary scientific integrals

Professor Matthew Schwartz has introduced BootLoops, an open-source toolkit designed to harness large language models for specific quantitative scientific tasks. Published in October 2026, this approach moves beyond general AI assistance by targeting "Claude-shaped" problems where agentic workflows excel at coding and mathematical orchestration rather than deep conceptual reasoning.

What happened

Schwartz initially attempted to use Claude Opus 4.5 as a standard research assistant but found the process inefficient. The model required constant correction and steering, creating an "impedance mismatch" between human scientific intuition and AI capabilities. Instead of forcing the AI to act like a scientist, Schwartz shifted strategy to identify problems that matched the current strengths of LLMs: vast knowledge breadth, strong coding skills, and rapid data parsing.

He began by tasking Claude Fable 5, released in Summer 2026, with porting methods from scattering amplitude research into a common framework. This effort resulted in BootLoops, a software harness that allows the AI to execute complex calculations. The toolkit quickly reproduced results that had taken Schwartz weeks to code in just twenty minutes. More importantly, it identified more efficient algorithms and extended these methods to elliptic integrals, a class of problems previously too difficult for complete bootstrap computation. Within weeks, the system computed thirty integrals end-to-end, including fifteen that had never been solved before.

The utility of BootLoops extended beyond high-energy physics. Because the underlying mathematical structures, such as Bayesian evidence integrals, appear in disparate fields, the toolkit found applications in ecology and population genetics. Schwartz collaborated with domain experts like ecologist James O’Dwyer and geneticist Michael Desai to validate these cross-disciplinary findings. In ecology, the tool solved a twenty-year-old equation from neutral biodiversity theory, revealing that tree species mix on Barro Colorado Island changes significantly faster than random chance predicts. In genetics, it addressed a thirty-year-old integral regarding natural selection and rare mutations using data from the gnomAD catalog.

How it works

BootLoops functions as a specialized harness for LLMs, similar to how Claude Code or Codex structure interactions for coding tasks. It targets the semi-numerical bootstrap method, which combines physical constraints with high-precision numerical computations to determine theoretical amplitudes. This approach is ideal for agentic AI because it requires integrating knowledge from mathematics, physics, and computer science while remaining strictly verifiable through numeric checks.

The system leverages the AI's ability to parse academic papers and existing codebases in languages like Wolfram Language, C++, Python, and Julia. By unifying these fragmented resources, BootLoops allows the model to generalize machinery from one domain, such as logarithmic functions in physics, to another, such as elliptic functions or ecological models. The AI handles the heavy lifting of code generation and algorithm optimization, while human experts provide direction on scientific relevance and interpret the results within their specific disciplinary contexts.

Key details

• BootLoops is an open-source toolkit compatible with various large language models, not limited to Anthropic’s offerings.
• The system reproduced complex physics results in approximately twenty minutes, a task that previously required weeks of manual coding.
• It successfully computed fifteen previously unsolved elliptic Feynman integrals by generalizing existing bootstrap methods.
• In ecology, the tool solved Stephen Hubbell’s neutral theory equation, showing species mix changes 4.5 times faster than predicted on Barro Colorado Island.
• Collaboration with experts was essential to steer technically correct but scientifically unremarkable AI outputs toward meaningful research questions.
• The approach relies on the "semi-numerical bootstrap," which uses high-precision numeric points to pin down theoretical coefficients exactly.

Why it matters

For software engineers and AI practitioners, this case study highlights the importance of designing workflows that match the actual capabilities of current models rather than their hypothetical potential. The "impedance mismatch" described by Schwartz is a common friction point in enterprise AI adoption. By identifying "Claude-shaped" problems—tasks that are well-scoped, code-heavy, and verifiable—developers can build more reliable agentic systems. This shifts the focus from trying to replace human expertise to augmenting it with high-speed computational orchestration.

Furthermore, the success of BootLoops demonstrates the value of cross-domain pattern recognition. AI agents can identify mathematical isomorphisms between unrelated fields, such as physics and biology, that human specialists might miss due to disciplinary silos. However, this also underscores the necessity of human-in-the-loop validation. Technical correctness does not equate to scientific significance, and domain experts remain critical for interpreting results and defining valuable problem statements. This hybrid model offers a practical blueprint for building AI tools in specialized industries.

What you can do

• Identify repetitive, code-intensive tasks in your workflow that require integrating multiple data sources or legacy codebases.
• Build lightweight harnesses that constrain LLM outputs to verifiable formats, such as unit tests or numeric checks, to reduce hallucinations.
• Collaborate with domain experts early in the development process to ensure the AI solves relevant problems rather than just technically feasible ones.
• Explore open-source tools like BootLoops to understand how structured prompts and tool orchestration can enhance model performance in specialized domains.
• Look for mathematical or structural similarities between problems in different departments or fields within your organization to leverage AI’s broad knowledge base.
• Treat AI agents as junior collaborators with specific strengths in coding and data parsing, rather than autonomous researchers capable of deep conceptual innovation.