FULL STORY

Claude Agents' Enzyme Discovery: From Hype to Skepticism

Anthropic announced its new wet lab's first result: ~950 Claude agents that spent 21 hours finding CRISPR-like enzyme systems. The claim drew buzz, Nature coverage, and skepticism from Copenhagen researchers.

2026-09-24 ~ 2026-09-30 · 4 episodes · 94 posts

Episode 1 · 950 Claude agents autonomously uncover novel CRISPR-like enzyme system in 21 hours (2026-09-24, 75 posts)

Anthropic's newly established molecular biology wet lab has delivered its first scientific discovery: roughly 950 Claude agents working in parallel for about 21 hours and consuming some 210 million tokens uncovered a previously unknown CRISPR-like enzyme system—Array-associated Reverse Transcriptases (ART)—in phage DNA, which CEO Dario Amodei called a possible new gene-editing mechanism. It is widely seen as a landmark case of agentic AI moving from concept to real scientific output, though the company acknowledges the system's function remains unclear and bioinformatics experts have questioned the finding's novelty and maturity.

Confirmed

  • Anthropic officially announced a new molecular biology research group and wet lab to explore whether Claude can help scientists spot anomalies in nature's molecular machines; team members have long worked on proteins, including tracing CRISPR evolution and hunting enzymes for cell and gene therapy.
  • Per The Verge (relayed by rohanpaulai), nearly 950 Claude agents worked for 21 hours and consumed about 210 million tokens searching massive DNA sequence databases; one agent noticed an anomalous repetitive DNA sequence that became the key lead. Finer statistics (rohanpaulai): 949 agents over 21.5 hours searched 1.94 billion protein clusters, recovered 198,290 reverse transcriptase clusters, and used 215.6 million tokens.
  • Per DrSingularity and others, while analyzing over 200,000 reverse transcriptases, Claude autonomously identified a previously uncharacterized enzyme system with CRISPR-like DNA repeats, named Array-associated Reverse Transcriptases: a known RT adjacent to another gene and a repetitive DNA array, forming a CRISPR-like programmable structure.
  • The team provided only an initial direction and downstream lab validation; most of the work—reading literature and genomic data, searching sequence databases, writing code, comparing sequences and genomic neighborhoods, and proposing hypotheses—was led by Claude (Amodei stressed "most," not all).
  • The company says the system's function is unclear, though the few known systems with similar features can all cut, copy and paste DNA; no paper or peer review details were disclosed.
  • Dario Amodei called it a possible new gene-editing mechanism whose exact function, biotech uses and importance remain to be confirmed, saying it is work he would have been proud of during his PhD; he also cited it as evidence of AI capability jumps, arguing biology is repeating the exponential progress seen in math since 2023.
  • Community discussion (1411337) compared it to OpenAI's 10,000-agent run last month as part of a larger agentic-science race; Polymarket opened prediction markets on the event.

Unconfirmed

  • The enzyme system's exact function, biotech applications and scientific significance are all unconfirmed—the company itself says it is unknown; no paper or peer review has been disclosed.
  • Bioinformatics experts (relayed via vishalmisra, rbhar90) consider the wet-lab results very preliminary and the methods not novel, noting the team knows little about how the system works and that the only real wet experiment so far was expressing the enzyme in E. coli; commenter zivravid (amplified by Gary Marcus) said the "AI-driven biological discovery" would not survive a PhD-defense-level review; Crescitaly and others asked whether it counts as a "discovery" at all, noting human scientists validated the lead afterward and the biological function remains undetermined.

Why it matters

  • It is the Anthropic wet lab's first result and an early demonstration of large-scale multi-agent systems autonomously producing new scientific knowledge, marking AI's shift from assistant to experimental workhorse.
  • The human-AI division of labor—Claude proposes hypotheses, scientists vet them, the wet lab validates—offers a reusable template for AI-accelerated basic biology; historically, discoveries of similar programmable systems such as CRISPR spawned an entire gene-editing industry, fueling high expectations.
  • Yet experts' caution suggests the paradigm itself—LLMs automatically mining anomalies in biological data for human labs to verify—may matter more than this specific finding.

55 more related posts →

Episode 2 · Anthropic's Wet Lab Debut: 950 Claude Agents' DNA Discovery Sparks Scientific Backlash (2026-09-24, 10 posts)

On September 23, Anthropic announced a molecular biology wet lab in San Francisco and its first result: roughly 950 Claude agents spent 21 hours (about 77 agent-hours) and 210 million tokens searching data and literature, finding a recurring DNA sequence pattern next to an unusual reverse transcriptase (RT) gene. Human Anthropic scientists reviewed the hypotheses and performed all wet-lab work; the company also opened biology-focused access and invited genomics collaborations. Bloomberg reported Anthropic comparing the enzyme system to a gene-editing breakthrough, triggering sustained scientific controversy over originality and hype.

Confirmed

  • Workflow: Claude generated hypotheses and candidate biological systems; human scientists reviewed and performed wet-lab validation.
  • Scale: 950 agents, 21 hours, 210 million tokens, 77 agent-hours.
  • Anthropic opened bio access and hired its own evaluators just 11 days after Dario Amodei's September 12 essay "We Must Pace the Frontier"; Sayer Ji and others questioned the company's stance.

Unconfirmed

  • Originality: biologist dvira noted Korn et al. reported a similar finding in 2021, and such bioinformatics predictions appear in dozens of papers daily—experimental validation is the scarce part; Anthropic has not responded.
  • soumitrashukla9 relayed Dvir's assessment that without PR the paper is fairly mundane bioinformatics; experts cited by Bloomberg also felt the significance was overstated, pending independent validation.

Why it matters

The core dispute is the gap between marketing framing and actual capability. dvira argued the AI autonomously did what a trained bioinformatics PhD could do given the same instructions, citing that a high school student in his pre-LLM lab found a synthetic-lethal anticancer target via data mining. eperlste countered: "A grad student could have done it with existing tools—so why did no one?" Sean McClure clarified the point: not that humans can't, but that general LLM agents may now handle a substantial portion of workflows previously requiring trained computational biologists. This cuts to how AI for Science should be judged—autonomous execution versus genuinely novel discovery—while self-hired evaluators deepen conflict-of-interest concerns.

Episode 3 · Anthropic's AI Wet Lab Scans 1.9 Billion Proteins, Finds CRISPR-like Viral DNA (2026-09-27, 3 posts)

Anthropic launched a biology wet lab where roughly 950 AI agents work with scientists, discovering CRISPR-like DNA patterns in giant viruses and screening 1.9 billion proteins to identify three candidate enzymes, though the findings await validation.

Episode 4 · Anthropic's Claimed AI Scientific Discovery of ARTs Enzymes Challenged by Researchers (2026-09-28, 6 posts)

Anthropic announced on September 23 the first result from its new biology lab: roughly 950 Claude agents running in parallel for 21.5 hours identified a CRISPR-like feature in massive DNA sequence data — new enzymes called ARTs (array-associated reverse transcriptases) with biotech potential — which the company framed as an autonomous AI scientific discovery. University of Copenhagen computational biologist Mario Rodríguez Mestre publicly disputed the claim, and the story gained wide attention after coverage by The New York Times, WIRED, and MIT Technology Review.

Confirmed

  • Anthropic officially announced the result, claiming AI agents discovered the ARTs enzymes; the announcement came September 23, with about 950 Claude agents running 21.5 hours across large genomic datasets.
  • Rodríguez Mestre said his team had studied these enzymes for four years, suggesting the claimed independent discovery may rest on prior human work, possibly leaked from unpublished research.
  • The New York Times, WIRED, and MIT Technology Review all covered the dispute. WIRED noted Anthropic issued PR before experimental validation; MIT Technology Review asked the deeper question of when AI can truly be said to have made a scientific discovery.
  • Piotr Binkowski commented "Nothing here was accidental" when resharing coverage, implying the finding was not the product of independent model reasoning.

Unconfirmed

  • Whether the ARTs discovery drew on or leaked from unpublished human work remains unsettled; the dispute is still developing.

Why it matters

  • The episode strikes at the credibility of "autonomous scientific discovery" claims in the AI industry: if a model's finding actually stems from human research already in its training data, such claims lose much of their value.
  • Releasing PR before experimental validation raises questions about AI companies' publication practices.
  • It is a landmark test of whether AI can genuinely advance science, touching on scientific credit and research integrity.