Harvard physicist Matthew Schwartz has released BootLoops, an open-source agent harness designed to help language models perform precise scientific calculations across unfamiliar disciplines. Used with Claude, the workflow contributed to 36 manuscripts in 18 fields over three months with 19 co-authors.
The harness organizes repeated cycles of proposing an approach, running tools, checking results and revising the work. That structure can bridge terminology and method gaps between subjects ranging from particle physics to ecology, linguistics and genome analysis.
The output count does not mean the system independently produced 36 validated discoveries. Schwartz says models frequently declare success too soon or draw an incorrect scientific conclusion from a calculation that is itself correct. Human specialists often turned an automated solution into useful research by identifying the meaningful question and checking every step.
BootLoops is therefore closer to a computational collaborator than an autonomous scientist. Its open code makes the process inspectable, but each manuscript still requires domain review, reproducible calculations and ordinary peer scrutiny. Schwartz’s practical instruction is direct: look at everything yourself. The tool may accelerate exploration across fields, while responsibility for deciding whether a result is novel, sound and relevant remains with researchers.