Verifiable Agentic Science on Public Infrastructure

Connecting autonomous coding agents to public computing infrastructure enables large-scale agentic reproduction and discovery with complete provenance, exposing modeling and reference mismatches across many studies.

• Why it matters: Agentic science risks irreproducible results when local improvisation leaves no verifiable record, undermining trust and repeatability.
• What they did: The work links coding agents to shared resources and first shows 30 unguided local agents splinter into 18 incompatible pipelines on the same dataset.
• The result: In a large-scale reproduction and discovery study, it reproduces a large, diverse array of studies with every workflow, history, and intermediate dataset publicly accessible, revealing unstated choices and subsampled tutorial inputs.