Verifiable Agentic Science on Public Infrastructure
Dannon Baker, Scott Cain, Marius van den Beek, John M. Chilton, Danielle Callan, Steven Weaver, Nate Coraor, Kelsey M. Beavers and 5 more
Why it has this license class
AChecked 2 Nov 2026. Open license (CC-BY, CC-BY-SA, CC0, public domain): full text indexed and used in synthesis.
| Source | License | Open-access status | Read as |
|---|---|---|---|
| openalex | cc-by | green | Green |
| crossref | http://creativecommons.org/licenses/by/4.0/ | — | Green |
| unpaywall | cc-by | green | Green |
Abstract
BAbstract Autonomous artificial intelligence agents are transforming the practice of science. Across every discipline, researchers are handing data analysis to coding agents that can design and execute complex studies from a single prompt. Yet this revolution carries a threat. When agents improvise in local environments, they leave no verifiable record, ultimately risking a wave of irreproducible science. Solving this crisis requires moving agentic science onto public infrastructure where every action leaves an examinable trail. Here we connect autonomous coding agents to shared computing resources, turning ephemeral prompts into auditable scientific records. After showing that 30 unguided local agents splinter into 18 incompatible pipelines on the same dataset, we conduct an agentic reproduction and discovery study on a large scale—reproducing a large, diverse array of studies with every workflow, history, and intermediate dataset publicly accessible. This complete provenance trail exposes unstated modeling choices, reference mismatches, and subsampled tutorial inputs in published studies, and establishes a new foundation where human scientists can understand, steer, and trust AI-driven discovery.
Claims built on this paper
D- Lu et al. 2026 warn that autonomous systems could tax review systems and add noise to the literature; Baker et al. 2026 ("Verifiable Agentic Science on Public Infrastructure") measure a concrete mechanism, showing that 30 unguided local agents given the same dataset splinter into 18 incompatible pipelines—an irreproducibility arising from unrecorded improvisation rather than from fabrication. Trace →
- Full-chain traceability is being built as shared infrastructure rather than only harvested from session transcripts: Baker et al. 2026 connect autonomous coding agents to shared computing resources and reproduce a large, diverse array of studies with every workflow, history and intermediate dataset publicly accessible. Trace →
- Agentic scrutiny of already-published human analyses is recurring across domains: Huynh et al. 2026's RepurAgent flagged confounders missed in manual review of a COVID-19 antiviral screen, and Baker et al. 2026's large-scale agentic reproduction exposed unstated modeling choices, reference mismatches and subsampled tutorial inputs in published studies. Trace →