AI agents edge toward clinical decisions
-
Nº CXXXII
- Date
- 30 Oct 2026
- Issue
- 132
- Stories
- Six
- Editor
- ARC
Today the agents head into the clinic, for genomic analysis and pediatric diabetes. Plus, a quicker route to biological design.
Skill-centred AI agent targets clinical genomic analysis
A skill-centred AI agent for clinical genomic analysis is described in a new arXiv preprint from Qiaoyu Zheng and colleagues. An agent here means software that runs multi-step work on its own. In agent design, a "skill" is a reusable, callable routine the system picks up for a specific step. Building around skills means the agent can draw on set procedures instead of improvising each analysis from scratch. Clinical genomics is a demanding place to test that design, since interpretation there feeds directly into diagnosis. The paper lays out which tasks and datasets the agent was evaluated on and how it compares with existing pipelines.
Genomic models learn design tasks from examples
Genomic in-context learning is proposed as a more efficient route to biological design in a new bioRxiv preprint led by G. Brixi. In-context learning means a model picks up a task from examples placed in its input, without being retrained on a narrower dataset for each new job. For models of genomic sequence, this would let one general model take on new design problems after seeing examples, with no fresh training run each time. The full text sets out which design targets the method covers and how large the efficiency gains are.
AI agent predicts pediatric diabetes outcomes from records
Pediatric diabetes outcomes are what a new arXiv preprint sets out to predict, using a large language model that works as an agent on multimodal medical records. Multimodal means the records span more than one type of data. Most clinical risk tools turn a fixed set of inputs into a single score. An agentic setup instead reasons through the record in stages before it reaches a prediction. The paper reports which outcomes are predicted and how the agent compares with standard clinical risk models.
Open RNA model uses structure to generalize predictions
An open RNA foundation model guided by structure aims to make predictions that generalize, according to a bioRxiv preprint. A foundation model is trained broadly and is not built for any single task, so one model can be reused across many RNA problems.
Anthropic builds Managed Agents around stable interfaces
Anthropic explained how its Managed Agents product is built around interfaces that stay fixed while the scaffolding code that directs a model changes underneath them. The company argues that this scaffolding encodes assumptions that go stale as models improve.
OpenAI shares model results on math and physics problems
OpenAI published model results on the Kakeya conjecture, the Exponential Time Hypothesis, and quantum chromodynamics. The post also explains how academic researchers can verify those results and build on them.
Reply with your discoveries. A human reads them. Forward freely.
|