Agentic Science and Physical AI: Claude’s Enzyme Search, Anthropic’s Wet Lab, FDA-Authorized Diagnostics, and What AI in Education, Science, and Medicine Now Requires
SCIENCE BIOMEDICINE AND PHYSICAL AI
AI moved from answering questions toward running research
The strongest science signal was organizational: hundreds of agents searched biological data while physical labs, clinical tools and autonomous fleets widened the action surface.
| Development | What is new | Proof still required |
|---|---|---|
| Claude enzyme search | Approximately 950 agents reportedly searched genomic data and identified a CRISPR-like candidate system. | Functional role, novelty, reproducibility and independent review. |
| Anthropic wet lab | The lab connects model planning to robotic experimentation. | Human oversight, sample identity, safe bounds, contamination control and rollback. |
| Sepsis ImmunoScore | Sinai Hospital reportedly began using an FDA-authorized AI diagnostic tool. | Clinical performance, workflow fit, bias monitoring and treatment impact. |
| Novo Nordisk and Claude Science | AI is moving deeper into literature analysis, study design and research workflows. | Traceable evidence, reviewer accountability and protected research data. |
| Autonomous vehicle fleet plan | Lucid and Bolt described a 25,000-vehicle European deployment plan. | Regulatory approvals, safety cases, incident evidence and operating economics. |
A physical-action control boundary
| Before action | During action | After action |
|---|---|---|
| Approved purpose; model and tool version; environmental limits; named human owner. | Authenticated device; least privilege; location and safety constraints; live stop path. | Observed outcome; immutable evidence; exception review; recovery; recurrence monitoring. |
Scientific generation is getting cheaper. Verification is becoming the scarce resource. The organizations that can preserve provenance, reproduce a result and explain the human acceptance decision will convert experimentation into trusted discovery.
INDUSTRY AND INSTITUTION WATCH PART FOUR
Education science and medicine
| Area | News and signal of interest |
|---|---|
| AI in Higher Education | A national system-head association launched an AI advisory board and readiness assessment; a provost survey found only 20 percent believed their institution had a coherent curricular vision. |
| AI in Education | AI4MiddleSchools expanded, EDSAFE widened its policy network and some parents called for pauses until independent safeguards exist. |
| AI in Biomedicine | Claude agents reported a preliminary enzyme-system discovery; Anthropic opened a biology lab; Novo Nordisk adopted Claude Science workflows. |
| NASA | No major announcement was identified in the reporting window. Recent work with IBM on a lunar foundation model and coordinated exploration AI remains the relevant signal. |
| National Institutes of Health | NIH emphasized practical adoption through its AI in Action program and training on advanced scientific workflows. |
Why the science coverage was different
The science stories did not claim that humans were removed. They showed AI organizing search at a scale that changes where human judgment is applied. Researchers increasingly set the objective, constrain tools, interpret candidates, run physical validation and decide whether evidence is sufficient. That makes research credit, provenance, reproducibility and acceptance criteria part of the system design.
A practical evidence chain
| Stage | Evidence to retain |
|---|---|
| Question | Research purpose, hypothesis, owner, prohibited paths and acceptance criteria. |
| Search | Data sources, query strategy, model version, agent roles, tokens, code and tool calls. |
| Candidate | Ranking logic, uncertainty, alternative explanations and human review. |
| Experiment | Protocol, instrument state, sample identity, deviations, observed results and safety stops. |
| Conclusion | Independent reproduction, reviewer decision, limitation and publication record. |