AI agents doing Nobel‑level work?
A March 27 roundup claims AI agents can compress complex physics analyses into 4–6 hours of agent‑assisted work and flags early uses in novel cancer immunotherapy research. (x.com) The post is part prognostic and part showcase — watch for reproducibility and tooling tightness if you plan to adopt agent pipelines. (x.com)
A new preprint dated March 23, 2026 from MIT and CERN researchers led by Eric A. Moreno reports an end‑to‑end agentic framework that autonomously handled event selection, background estimation, uncertainty quantification, statistical inference and draft paper writing in experimental high‑energy physics. (arxiv.org) The authors say they reproduced eight historical analyses using archived ALEPH, DELPHI and CMS open datasets and ran the pipeline on Anthropic’s Claude Code backend using the claude‑opus‑4‑6 model variant. (arxiv.org) Multiple secondary writeups and a video walkthrough describe the agent-driven reproductions completing in roughly hour‑scale runs — several summaries quote an approximate six‑hour wall‑clock for full agent execution on selected cases. (emergentmind.com, emergentmind.com) In parallel, oncology groups are publishing early agentic work: a multi‑phase medRxiv preprint from Mass General Brigham, Harvard and collaborators tests an agentic system for clinical detection of immunotherapy toxicities, and an arXiv preprint titled “Bio AI Agent” outlines a multi‑agent pipeline for autonomous CAR‑T target discovery and molecular design. (medrxiv.org, arxiv.org) Community commentary and reviews warn about verification and governance: Nature’s October 3, 2025 guide urged explicit validation standards for agentic research and flagged risks from narrow scaffolding and error propagation, while an Institute of Physics community perspective highlights training, reproducibility and V&V challenges in physics. (nature.com, iop.org) The MIT/CERN paper itself calls for concrete changes — new training programs, reorganized analysis teams, and built‑in multi‑agent review steps via its “Just Furnish Context” framework to try to catch propagation errors — so watch for independent reproductions, open toolchains and provenance records as the next validation steps. (arxiv.org, arxiv.org)
