Anthropic’s Claude Discovers a Novel Enzyme System by Searching Biology at Agent Scale
Anthropic says Claude found a CRISPR-like enzyme system in DNA data, illustrating how AI agents can search biology while experiments remain the proof.
A biological discovery rarely begins with a clean label. CRISPR was first noticed as an unusual repeat pattern in bacterial DNA; Taq polymerase emerged from a Yellowstone microbe. Anthropic says Claude has now identified a previously unrecognized enzyme system associated with repeating DNA sequences, after roughly 950 agents searched a large database for 21 hours and used 210 million tokens. The finding is intriguing, but the scientific story is not that a chatbot invented a gene-editing tool. It is that a model generated a candidate that still has to survive laboratory evidence. Anthropic’s research report is the primary source for the agent count, token count, reverse-transcriptase context, and the limits of the current finding.
The evidence behind the story
| Reporting boundary | What the primary source says | What still needs independent testing |
|---|---|---|
| Central claim | Anthropic’s Claude Discovers a Novel Enzyme System by Searching Biology at Agent Scale | Performance and impact on a reader's workload |
| Evidence | Named announcements, documentation, or research | Replication, operational data, and failure cases |
flowchart LR
A[Primary source] --> B[Technical claim]
B --> C[Independent test]
C --> D[Operational decision]
D --> E[Observed result]
E --> C
The discovery began as a search problem
Anthropic’s September 23, 2026 report describes an early result from a new life-sciences research group and laboratory. The company says Claude autonomously discovered a novel enzyme system associated with an array of DNA repeats, a pattern reminiscent of CRISPR. The function is not yet known. That sentence should remain at the center of the story. A sequence pattern and a promising candidate are not a working biological tool, and a model’s observation is not a validated mechanism.
Why a reverse transcriptase matters
The project starts from a familiar scientific problem: biological databases contain more sequence diversity than a human team can inspect manually. Researchers can ask a model to search for uncharacterized protein families, compare motifs, propose candidates, and prioritize experiments. Anthropic says the work is part of a broader attempt to make agents collaborate with scientists at each stage. The interesting shift is not replacing the scientist. It is increasing the number of hypotheses that a scientist can decide whether to test.
CRISPR-like does not mean CRISPR
The reported system is based on a reverse transcriptase found in a jumbo phage. Reverse transcriptases copy RNA into DNA and are already a known class of enzymes. Anthropic says the underlying reverse transcriptase had been identified in previous studies, while Claude appears to have noticed a combination of features: an associated array of non-coding DNA sequences and an additional accessory protein of unknown function. The novelty lies in the relationship between the parts, not in discovering that reverse transcriptases exist.
Nine hundred fifty agents change the search economics
CRISPR-like is a careful phrase. CRISPR systems can provide programmable operations such as cutting, copying, or pasting DNA, but a repeated sequence near a reverse transcriptase does not prove that the new system performs any of those operations. The label is a clue about structural resemblance and research relevance. It should not be turned into a claim that the system is ready for gene editing, medicine, or industrial biotechnology.
The laboratory remains the authority
Anthropic reports that roughly 950 agents searched the data for 21 hours and consumed 210 million tokens before one agent found the striking pattern. Those numbers show the scale of the search, but they do not by themselves establish efficiency. Token use is not a scientific output, and many agents may have explored redundant paths. A meaningful comparison would measure candidate quality, search coverage, false leads, analyst time, and laboratory hit rate against a conventional pipeline.
What Claude noticed in the sequence data
Multi-agent search can be useful when the hypothesis space is broad and the data has structure. One agent can identify families, another can compare genomic neighborhoods, another can check known literature, and a later stage can rank candidates. The danger is correlated error: agents may repeat a plausible interpretation because they share model priors or because an early summary influences every later search. Independent branches, provenance, and negative results are necessary to tell exploration from consensus theater.
Human direction was limited but not absent
Anthropic says the scientists provided high-level direction and carried out the laboratory work, while the agents investigated reverse-transcriptase families and selected interesting candidates. That division is significant. Humans defined the research question, supplied the environment, and controlled experiments. The agents expanded the search and made judgments within the task. “Autonomous” therefore describes a portion of the workflow, not a lab without scientists, instruments, safety review, or experimental interpretation.
A candidate can be interesting without being useful
The laboratory is where the candidate becomes falsifiable. Researchers must express the relevant proteins, test whether the accessory component changes activity, characterize substrate preferences, and determine whether the repeat array has a functional role. A computational association can disappear under experimental conditions. It can also lead to a different discovery if the original interpretation is wrong. The correct response to a model-generated hypothesis is not belief; it is a test designed to fail cleanly.
The safety boundary is built into the uncertainty
Feng Zhang, a CRISPR pioneer and professor at MIT and the Broad Institute, described the identification as intriguing and worth further investigation, according to Anthropic’s report. That is an expert reaction to a research direction, not an independent validation of the system’s function. Attribution matters because vendor announcements can accidentally turn cautious scientific interest into a headline about confirmation. Readers should preserve the difference between “merits study” and “works.”
How to reproduce an agentic biology claim
The examples of restriction enzymes, Taq polymerase, and CRISPR in Anthropic’s account explain why unusual biological patterns matter. Restriction enzymes became useful tools after researchers recognized bacterial defense mechanisms. Taq polymerase enabled PCR after its heat tolerance was understood. CRISPR moved from a repeated sequence observation to a transformative editing platform through years of work. Those histories are reminders that discovery is a chain: observation, mechanism, reproducibility, engineering, safety, and application.
The evidence chain from database to experiment
A reproducible agentic biology report needs more than the final sequence. It should identify the database version, search criteria, prompts or task specifications, agent models, tool permissions, intermediate candidates, and the rule used to select the winner. Without that record, another team cannot distinguish a rare insight from a lucky post hoc explanation. Model updates make this harder, so the research environment should preserve versions and hashes just as a computational experiment would.
What this means for scientists with small teams
Provenance is especially important when agents consult literature or public databases. A model can misread an annotation, merge two papers, or treat an unreviewed claim as established. Each candidate should carry links to the original sequence record, relevant studies, alignment outputs, and the exact reasoning that led to the experiment. The record does not need to expose every internal model token, but it must expose enough evidence for a scientist to audit the decision.
The data problem is as important as the model
The workflow could help small research groups by reducing the first-pass search burden. A lab with a narrow domain expert may not have the capacity to examine every reverse-transcriptase family across a massive database. An agent can generate a ranked queue, summarize anomalies, and prepare a test plan. That advantage is strongest when the lab owns the experimental validation and when the system is allowed to return “not enough evidence” rather than being rewarded for a novel story.
Why autonomous research needs provenance
The data quality problem is unavoidable. Sequence databases contain incomplete annotations, sampling bias, contamination, duplicated records, and naming conventions that reflect past assumptions. An agent trained on published descriptions may inherit those assumptions and overlook an important pattern outside well-studied organisms. Search systems should expose coverage and uncertainty, not present a candidate ranking as a map of biological truth.
The promise and limits of high-throughput hypotheses
Safety has two dimensions here. First, an unknown enzyme may have properties that require careful containment and review. Second, an agent capable of searching and designing biology workflows could be misused if tool access crosses from analysis into unrestricted procurement or experimental execution. Anthropic’s report emphasizes discovery and laboratory work, but the broader lesson is to separate computational hypothesis generation from physical action with authorization, screening, and expert oversight.
Biology may be where agents learn humility
An effective evaluation would compare Claude’s agent team with expert scientists and simpler computational baselines on a blinded set of biological discovery tasks. Score novelty, correctness, experimental tractability, time to candidate, and the proportion of suggestions that fail for obvious reasons. A system that generates ten times as many hypotheses but burdens scientists with ten times as much noise may not improve research. The measure is useful discoveries per unit of expert attention.
Operational questions for the teams adopting this work
Sequence novelty is not functional novelty
A candidate can be new to a database because it was missed, poorly annotated, or simply different from known examples. Functional novelty requires assays and controls. Researchers should test whether the observed protein is expressed, stable, active, and specific under defined conditions. The agent can prioritize the sequence, but the experiment must determine whether the biological behavior exists.
The repeat array is a hypothesis anchor
The associated non-coding DNA repeats are interesting because they may be part of a system rather than decoration. Researchers can compare repeat length, spacing, conservation, and genomic context across related organisms. They can also test whether changing the array changes activity. Those experiments turn a visual pattern into a causal question. Until then, the repeats remain evidence for attention, not proof of mechanism.
Jumbo phages widen the search space
Large bacteriophages carry unusual genetic material and can be underrepresented in reference collections. Searching them may expose systems that standard bacterial annotations overlook. It also increases the need for careful contamination checks and lineage analysis. An agent that searches broadly is valuable only if it can distinguish an authentic biological signal from a database artifact or assembly error.
Agent ensembles need disagreement
If every agent shares one ranking, the result may look confident without being independent. A stronger design asks different agents to use different search strategies, then records where they disagree. Human scientists can inspect candidates with diverse rationales and prioritize tests that separate competing explanations. Disagreement is information about uncertainty, not noise to remove from the report.
Lab automation changes the risk profile
Once a computational system can order materials, schedule instruments, or run experiments, the boundary between suggestion and action becomes important. High-level design can be automated more freely than physical execution. Procurement screening, biosafety review, access control, and human authorization should remain explicit. Faster hypothesis generation should not silently expand the set of experiments a system may perform.
Negative results are a training asset
A candidate that fails in the lab can teach the search system what a false visual pattern looks like. If teams record failed assays and the reasons for rejection, later agents can rank more intelligently. Publishing only successful discoveries creates survivorship bias and makes the method look more magical than it is. The negative queue is part of the scientific dataset.
The model needs domain tools
A language model alone cannot inspect every sequence reliably. Aligners, structure predictors, phylogenetic tools, database APIs, and literature search are the instruments around the agent. Their versions and settings affect the result. A research record should identify those tools so another lab can reproduce the path rather than trusting a prose summary of what the agent supposedly did.
Safety review should start at discovery
A sequence that looks like a useful biological system may have implications before its function is fully understood. The lab should classify candidates, restrict dissemination of sensitive operational detail where appropriate, and involve biosafety experts. Responsible disclosure is not a postscript after a model has found something interesting. It is part of deciding what to test and how to share it.
Credit must include the infrastructure
The final paper may name the model and the scientists, but the discovery also depends on databases, compute, lab technicians, protocols, and reviewers. Recording those contributions improves reproducibility and avoids treating agent scale as the sole source of insight. A well-documented team can show which steps were automated and which required human judgment.
The benchmark should be discovery yield
Token count and agent count are easy to report, but the valuable metric is experimentally validated findings per unit of scientist time and compute. Track candidate triage, assay success, time to result, and the cost of dead ends. If the system finds one strong candidate among a thousand noisy suggestions, the workflow may still be useful—but only if the lab can process the queue.
A discovery system must preserve surprise
If the search prompt defines only known families, an agent may optimize for familiar patterns and miss the unusual system that makes discovery valuable. Scientists should allow anomaly queues and review candidates that score poorly on ordinary similarity but strongly on structural evidence. The point of scale is not to automate the existing literature’s blind spots. It is to inspect the parts of biology that standard labels leave behind.
1. A sequence candidate should carry a confidence interval or competing explanations rather than one definitive label.
If the search prompt defines only known families, an agent may optimize for familiar patterns and miss the unusual system that makes discovery valuable. Scientists should allow anomaly queues and review candidates that score poorly on ordinary similarity but strongly on structural evidence. The point of scale is not to automate the existing literature’s blind spots. It is to inspect the parts of biology that standard labels leave behind.
2. Scientists need interfaces that expose the raw alignment and genomic neighborhood behind an agent summary.
If the search prompt defines only known families, an agent may optimize for familiar patterns and miss the unusual system that makes discovery valuable. Scientists should allow anomaly queues and review candidates that score poorly on ordinary similarity but strongly on structural evidence. The point of scale is not to automate the existing literature’s blind spots. It is to inspect the parts of biology that standard labels leave behind.
3. The laboratory protocol should be preregistered when practical so a compelling model narrative does not move the goalposts.
If the search prompt defines only known families, an agent may optimize for familiar patterns and miss the unusual system that makes discovery valuable. Scientists should allow anomaly queues and review candidates that score poorly on ordinary similarity but strongly on structural evidence. The point of scale is not to automate the existing literature’s blind spots. It is to inspect the parts of biology that standard labels leave behind.
4. A useful agent can identify the cheapest experiment that separates two hypotheses.
If the search prompt defines only known families, an agent may optimize for familiar patterns and miss the unusual system that makes discovery valuable. Scientists should allow anomaly queues and review candidates that score poorly on ordinary similarity but strongly on structural evidence. The point of scale is not to automate the existing literature’s blind spots. It is to inspect the parts of biology that standard labels leave behind.
5. Cross-lab replication is especially important when the discovery originated inside a vendor’s private system.
If the search prompt defines only known families, an agent may optimize for familiar patterns and miss the unusual system that makes discovery valuable. Scientists should allow anomaly queues and review candidates that score poorly on ordinary similarity but strongly on structural evidence. The point of scale is not to automate the existing literature’s blind spots. It is to inspect the parts of biology that standard labels leave behind.
6. The research group should publish enough negative evidence to show that the search was not curated after the fact.
If the search prompt defines only known families, an agent may optimize for familiar patterns and miss the unusual system that makes discovery valuable. Scientists should allow anomaly queues and review candidates that score poorly on ordinary similarity but strongly on structural evidence. The point of scale is not to automate the existing literature’s blind spots. It is to inspect the parts of biology that standard labels leave behind.
Deployment note 1
A useful interface should let a scientist reject a candidate with a reason that the next search can use. “Not novel,” “not expressible,” “unsafe to test,” and “duplicate of known family” are different labels. Treating every rejection as one negative signal would teach the system the wrong lesson and reduce future search quality.
Deployment note 2
The discovery process also depends on where agents stop. A search that explores every branch can consume enormous compute without improving the candidate set. Researchers need stopping criteria based on novelty, evidence, and experimental capacity. The best autonomous system may be the one that knows when its next ten thousand searches are less valuable than one carefully designed assay.
Deployment note 3
A paper or technical report should separate the agent’s discovery from the laboratory’s interpretation. That makes later corrections easier and protects the scientific record if the function turns out to differ from the initial hypothesis. Clear boundaries are not a public-relations weakness; they are how a surprising result remains useful after the first story changes.
Deployment note 4
The long-term opportunity is a shared infrastructure for hypothesis queues. Labs could contribute failed candidates, assay outcomes, and domain annotations without handing every proprietary dataset to one model provider. Such a system would make AI-assisted biology more cumulative and less dependent on a single company’s private agent traces.
Deployment note 1
A useful interface should let a scientist reject a candidate with a reason that the next search can use. “Not novel,” “not expressible,” “unsafe to test,” and “duplicate of known family” are different labels. Treating every rejection as one negative signal would teach the system the wrong lesson and reduce future search quality.
Deployment note 2
The discovery process also depends on where agents stop. A search that explores every branch can consume enormous compute without improving the candidate set. Researchers need stopping criteria based on novelty, evidence, and experimental capacity. The best autonomous system may be the one that knows when its next ten thousand searches are less valuable than one carefully designed assay.
Deployment note 3
A paper or technical report should separate the agent’s discovery from the laboratory’s interpretation. That makes later corrections easier and protects the scientific record if the function turns out to differ from the initial hypothesis. Clear boundaries are not a public-relations weakness; they are how a surprising result remains useful after the first story changes.
Deployment note 4
The long-term opportunity is a shared infrastructure for hypothesis queues. Labs could contribute failed candidates, assay outcomes, and domain annotations without handing every proprietary dataset to one model provider. Such a system would make AI-assisted biology more cumulative and less dependent on a single company’s private agent traces.
Sources readers can inspect
The links below are direct primary or first-party references used for the factual frame. Vendor claims are identified as claims in the article; interpretation and recommendations are editorial analysis.
- Anthropic novel enzyme report
- Anthropic Life Sciences Verification
- NCBI GenBank
- NCBI BLAST
- National Human Genome Research Institute CRISPR
- Nature CRISPR history
- NIH reverse transcriptase
- EMBL-EBI training
- Feng Zhang lab
- NIST AI risk management
The source record matters here. Vendor announcements establish what was built and what the publisher claims. They do not replace independent testing on a reader’s workload, so the practical recommendation is to preserve the distinction between a reported capability, an engineering inference, and a result that still needs measurement.