Loading page…
The model-independent toolkit packages methods developed with Claude, but Schwartz’s research examples show why correct calculations still need scientific judgment.
Listen to this story
BootLoops turns methods Schwartz developed with Claude into an open-source harness that other models can use for exact scientific calculations, making the workflow reusable beyond the original Claude sessions. It pairs research protocols with software and checks, but does not remove the need for domain expertise or human scrutiny. Schwartz reports promising physics and ecology results; several projects remain under exploration or verification, and Anthropic’s funding and publication of the work do not make BootLoops an officially supported Anthropic product.
Schwartz says Claude reproduced a paper’s results in about 20 minutes, versus weeks spent writing the original code.
BootLoops reportedly reproduced 15 known integral results and produced 15 additional results that had not previously been computed.
In a Barro Colorado Island forest analysis, the calculated rate of tree-species composition change was 4.5 times the rate allowed by neutral theory.
Theoretical physicist Matthew Schwartz released BootLoops 1.0 on October 1, 2026, putting tools developed with Claude into an open-source package for exact scientific calculations. Released under the MIT License, the toolkit combines software and research protocols around an AI model rather than treating a chatbot as a scientist.
BootLoops is a harness: a layer of tools and procedures that helps a model carry out a particular kind of work. In his guest post published by Anthropic, Schwartz says it can work with different models. The release makes that layer reusable beyond the Claude sessions in which he developed it.
The project began with scattering amplitudes, calculations that connect particle-collision data to the particles involved. Some of the underlying integrals can occupy a research group for years. Schwartz asked Claude to combine methods from his papers and nearby research into a common software framework.
One approach, called the bootstrap, uses physical constraints to narrow possible answers. A related method adds numerical calculations at extremely high precision to determine the remaining coefficients exactly. That combination suited AI-assisted coding: it drew on techniques scattered across papers and programming languages, while allowing answers to be checked against the original integral.
Schwartz reports that Claude reproduced one paper’s results in about 20 minutes, compared with weeks he spent writing his original code. He then pushed it beyond logarithms into harder elliptic functions. He says Claude generalized the existing machinery and wrote most of the additional software itself.
Schwartz says BootLoops reproduced 15 known integral results using the new method.
Schwartz reports 15 additional integrals that had not previously been computed, bringing the total to 30.
The same mathematical forms appear in different disciplines, giving the toolkit applications beyond physics. But Schwartz says Claude’s cross-field suggestions were often technically correct without being scientifically interesting. Domain experts helped turn those calculations toward questions their fields actually cared about.
His ecology example makes the distinction concrete. Schwartz says Claude solved a biodiversity equation that had resisted calculation at scale for 20 years. Applied to forest data from Barro Colorado Island in Panama, it showed tree-species composition changing 4.5 times faster than neutral theory allowed. That theory asks how much ecological change can be explained by random chance.
Plant biologist James O’Dwyer told Schwartz that ecologists already understood, more qualitatively, that neutral theory could not keep pace with real forests. He proposed subtracting its prediction and studying what remained. Schwartz says their resulting model of species life histories closely matched data; they are extending it to other forest plots.
Schwartz’s setup ran separate Claude Code sessions on Google Cloud virtual machines connected to GitHub and Overleaf. A master session coordinated projects, allocated computing resources and validated results. Background agents saved intermediate work in markdown files.
He also describes recurring failures: premature declarations of success, unreliable time estimates, inefficient multiday calculations and lost context in long sessions. His responses included putting protocols into the harness, inspecting plots himself and rechecking work as an adversarial referee. Those checks were part of the research process, not a substitute for it.
The scientific results are Schwartz’s account of work with collaborators. Several projects described in the Anthropic post remain under further exploration and verification. The release therefore provides tools to examine and extend, while those research claims retain their own verification status.
Unite.AI’s release details describe 49 tool packages across six repositories. The repository credits Schwartz as creator and Claude as the code’s writer under his supervision. It says Schwartz maintains the release and that it is not an officially supported Anthropic product.
Anthropic funded the project, and Schwartz worked there as a visiting researcher. Anthropic identifies BootLoops as owned and maintained by Schwartz, not as an Anthropic project. Publication on the company’s research site should not be confused with an official product-support commitment.
Loading discussion...
Join the conversation
Explain when an easier answer is worth changing the research question.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.