What happened
On October 1, 2026, theoretical physicist Matthew Schwartz released BootLoops 1.0, an open-source toolkit designed to facilitate exact calculations across various scientific disciplines using large language models. Developed during a three-month period, the project utilizes a harness that coordinates LLM sessions—specifically Claude Fable 5—to perform complex mathematical tasks, including the computation of Feynman integrals, population genetics modeling, and economic data replication. The toolkit is released under the MIT License and is maintained by Schwartz, with Anthropic providing project funding.
BootLoops 1.0 was released publicly on October 1, 2026, following a three-month development phase. The toolkit functions as a harness that manages Claude Code sessions on Google Cloud virtual machines, linking to GitHub and Overleaf repositories to coordinate research tasks. It is written in Python 3.12 with some Julia components and has been validated on Linux x8664 and Debian containers.
The toolkit was used to produce 36 manuscripts across 18 fields, including particle physics, ecology, population genetics, and economics. Notable achievements include the computation of 15 previously unsolved elliptic Feynman integrals and the replication of thousands of economics papers, where the workflow identified discrepancies and optimized calculation runtimes.
Schwartz reports that the system uses a master session to allocate and validate results, while background subagents store intermediate data. The project is maintained by Schwartz, though the repository notes that copyright for the code is held by Anthropic PBC. The toolkit is explicitly designated for research use and is not intended for clinical, regulatory, or public-safety applications.
Why it matters
BootLoops 1.0 represents a significant attempt to standardize the use of LLMs for high- scientific research, moving beyond general-purpose prompting toward a structured, verifiable workflow. By automating the translation of legacy code and the execution of complex mathematical integrals, the toolkit aims to accelerate discovery in fields ranging from particle physics to ecology. The project’s documented success in reproducing established results and solving previously intractable problems suggests a practical path for integrating AI into rigorous academic research, provided that human oversight remains central to the validation process.
The project demonstrates a shift toward 'Claude-shaped' problems—tasks that align with the strengths of current LLMs in symbolic manipulation and code generation. By providing a structured harness, Schwartz addresses common LLM failure modes, such as context loss and premature completion, through encoded protocol skills and adversarial human review.
The collaboration with NBER researchers to audit 4,452 economics papers highlights the potential for AI to improve the reproducibility of scientific literature. By porting legacy code from commercial tools like Stata and MATLAB into open-source Python, the project also promotes greater transparency in computational research.
The toolkit's ability to solve a 30-year-old integral in population genetics and apply it to large-scale genomic datasets like gnomAD underscores the potential for AI to act as a force multiplier for individual researchers, allowing them to tackle problems that would otherwise require years of manual coding.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
crm_get_transaction(id='4092').In AI, what are a model's "parameters"?
What to watch next
The primary focus is whether the scientific community adopts the BootLoops framework for peer-reviewed research and if the reported accuracy holds up under independent verification. Users should monitor the toolkit's performance across different LLM backends, as Schwartz notes the harness is designed to be model-agnostic. Additionally, the project's limitations—such as the model's tendency to declare premature victory or lose context in long sessions—highlight the ongoing necessity for expert human intervention in AI-assisted scientific workflows.
The project site states that BootLoops is model-agnostic, meaning it can theoretically be used with Gemini or ChatGPT. Future updates may reveal how performance varies across these different architectures when applied to the same scientific benchmarks.
Schwartz documented recurring failure modes, including the model's tendency to provide inaccurate time estimates and its struggle with long-session context. Future iterations of the toolkit will likely need to address these reliability issues to become a standard tool for the broader scientific community.
The project's reliance on human oversight—specifically Schwartz's role as an 'adversarial referee'—remains a critical component. It remains to be seen if the toolkit can maintain its accuracy levels when used by researchers who may not possess the same level of domain expertise to verify the model's outputs.