DiscoveryAgent pairs autonomous AI research with scientist steering and audit agents
Researchers describe a domain-agnostic framework where scientists redirect hypotheses between rounds while audit agents check checkpointed records, reporting strong results on two predictive tasks.
In plain words
A team has built DiscoveryAgent, a system where AI agents carry out research steps on their own, but scientists can inspect, change, or replace the hypotheses guiding each round. Separate audit agents review saved records of the work to trace claims and reproduce analyses. In tests, the system produced strong predictive models while keeping researchers in control of the research direction.
What was announced
The authors frame the core design problem as balancing autonomous investigation with scientists' ability to inspect and redirect ongoing work. Their answer combines autonomous research with scientist-directed hypothesis steering: researchers select, refine, or replace hypotheses between rounds, while audit agents interrogate checkpointed records to trace claims, reproduce analyses, and inform later decisions. The team describes this checkpointing system as a scalable form of oversight for agentic research.
On performance, the paper reports that DiscoveryAgent produced the best variable-star classifier among evaluated baselines, and, with scientist steering, a protein variant-effect predictor surpassed by only one of 106 ProteinGym baselines. Audit agents autonomously identified errors at higher rates than previous AI scientist agents, and scientists working with them found 14 times more valid errors than scientists doing manual review. Three expert-led case studies followed: a multimodal photometric-redshift estimator reaching 0.971 AUROC for low-redshift classification, documentation of an 83% rise in Florida homeowners-insurance premiums from 2013 to 2023 with a growing regressive burden, and identification of CD4+ T-cell relay genes enriched for autoimmune disease associations. The authors say the results show autonomous research expanding scientists' empirical reach while centering their authority over research direction.
What it does not show
- The source is the paper's abstract only; methods, evaluation details, and peer review status are not covered, and the reported results have not been independently confirmed.
For security and IT teams
- Scope
- The published evidence is limited to the paper's abstract, covering two predictive modeling tasks and three case studies rather than full methodology.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.