950 AI Agents, 21 Hours, One Discovery: What Claude’s CRISPR-Like Find Means for Your Business

Published 25 September 2026

The short answer: the AI story everyone has been talking about this week is not a new model or a price cut. It is a discovery. On 23 September 2026, Anthropic unveiled a new life sciences lab and said its Claude agents had found a previously undescribed enzyme system hidden in the DNA of bacteriophages, the viruses that infect bacteria. Roughly 950 agents searched a DNA database for about 21 hours, used around 210 million tokens, and flagged a pattern of repeating DNA that resembles the arrays behind CRISPR gene editing. Anthropic calls the system ART, short for array-associated reverse transcriptases. Nobody yet knows what it does.

The biology is fascinating, but the bigger trend is the method. This is one of the clearest public examples so far of an AI agent swarm doing open-ended search, filtering its own ideas and handing humans a short list worth their time. That same pattern works on contracts, security logs, codebases and customer feedback. Here is what happened, what it does and does not prove, and how a business can run its own small version safely.

Key takeaways

Anthropic says Claude agents gathered more than 200,000 reverse transcriptase enzymes, picked out 3,500 new candidate systems and narrowed them to 20 compelling candidates with written reports. One of those led to ART.
Humans set the question and did all the lab work. The agents did the searching, checking and write-ups, work Anthropic says would take an expert weeks to months.
The finding is early: the preprint is not peer reviewed, the system’s function is unknown, and Anthropic’s CEO acknowledged that a Stanford team had previously found a system that is similar in some ways.
The business lesson is the funnel: cheap, massively parallel agent search followed by expensive human judgment. Your reviewers, not your AI budget, become the bottleneck.
Running a swarm on your own data brings real risks: cost blowouts, confident but wrong reports, data exposure and dual-use concerns. Guardrails come first.

Illustration of a DNA double helix with one section of evenly spaced repeats highlighted in amber and a marker above it, surrounded by small dots representing AI agents — Hundreds of agents searched the data. One noticed an evenly spaced repeat pattern that no one had described.

What Anthropic actually announced

Anthropic formed a molecular biology research group in spring 2026 to test whether general AI models can speed up the kind of discoveries that launched modern biotechnology, from restriction enzymes to PCR to CRISPR. Its lab in the Bay Area looks like an ordinary molecular biology lab. The company says it works only at the lowest biosafety levels (BSL-1 and BSL-2), handles no pathogens that can infect humans, and that all lab work is performed by human scientists.

For the ART project, the team gave Claude a single high-level prompt: search a massive DNA database for interesting new examples of reverse transcriptases, the enzymes that copy RNA into DNA. From there, the agents surveyed the enzyme families on their own, reproduced known results to check their methods, and judged which candidates were worth a closer look. While reading raw DNA next to one unusual enzyme, an agent noticed a long run of evenly spaced repeats, counted and measured them, compared the layout with known systems, searched the literature for earlier reports, and filed a report for human review.

ART has three parts: the enzyme, a neighbouring partner gene of unknown function, and the repeat array. The underlying enzyme, found in a so-called jumbo phage, had been identified in earlier studies. What Anthropic says is new is noticing the surrounding features that mark it as a distinct system. Its first experiments found that the array is expressed as a set of distinct short RNAs, which hints at a CRISPR-like mechanism but does not prove one. CRISPR pioneer Feng Zhang of MIT and the Broad Institute reviewed the preprint and called the finding intriguing and worth further investigation.

How 950 agents turned 200,000 candidates into one lead

Stage	How many	Who did the work
Enzymes gathered from the DNA database	200,000+	Claude agents
New candidate systems picked out	3,500	Claude agents
Compelling candidates written up as reports	20	Claude agents, self-reviewed
Lead taken into the lab	1 (ART)	Human scientists

The ART discovery funnel: about 950 agents, about 21 hours and about 210 million tokens, according to Anthropic. Each stage is cheaper for software than for people, until the last one.

The numbers matter because they show where AI changed the economics. Anthropic says this kind of genome-mining survey can take an expert scientist weeks to months. The agents compressed it into less than a day by running in parallel, then did something just as important: they eliminated most of their own ideas before a human ever saw them. Anthropic says most candidates are typically cut at the self-review stage, and a survey may end with a single candidate worth testing, or with none.

The team also made an unusual admission. Claude produces hypotheses so prolifically that the hypotheses themselves have become an object of study. With hundreds to thousands of candidate reports per campaign, the scientists now analyze which proposals they choose to test and feed those lessons back into Claude’s instructions, teaching it their scientific taste.

What the discovery does and does not prove

Healthy skepticism is warranted, and Anthropic has been fairly open about the limits.

The function is unknown. ART shares a combination of features found in only a handful of other systems, all of which can cut, copy or paste DNA. A similar layout is a reason to investigate, not evidence that ART can edit genes.
It is not peer reviewed. The results were shared in a blog post and a technical preprint. Validation is up to the wider research community.
It was hard to repeat. According to The Next Web’s reading of the preprint, Anthropic reran the same search ten more times and every rerun missed the repeat array, because the agents did not read far enough into the surrounding DNA. The team attributes this to the size of the search and the agents’ unpredictable behaviour. For a business, that is the key caution: a swarm that finds something once may not find it again, so one clean run is not proof that nothing is there.
It builds on others’ work. CEO Dario Amodei wrote that the discovery was made mostly, though not entirely, by Claude, and acknowledged a Stanford team had earlier found a system that is similar in some ways.
Critics question the timing. Some AI researchers argued on social media that announcing a discovery before knowing what it does invites hype. Anthropic says it shared early to show what Claude can do and to open the work to other scientists.
AI in biology is not new. Google DeepMind’s AlphaFold, Stanford’s work pairing language models with CRISPR, and UCSF’s AI-designed enzymes all predate this. What is new is a general-purpose model autonomously driving the search end to end.

Why this is the AI trend to watch

For three years, the business conversation about AI has been about answers: chatbots, copilots and summaries. September 2026 has been about agents that act, from personal agents that spend money to agents replacing the app screen. The ART result adds a third category: agents that search for things nobody knew to ask about.

That shift is only practical because running agents got cheap. As we covered in The 90-Minute AI Price War, Anthropic and OpenAI cut frontier prices on 22 September. Anthropic did not publish what the ART search cost, and press coverage of the preprint suggests it used a model more capable than the public ones, but the token count still lets you bracket the cost if a business ran a similar job on today’s public models.

Scenario	Assumption for 210M tokens	Estimated model cost
Cache-heavy agent work	Mostly cached reads on Claude Opus 5.5 at $0.20 per million	About $40 to $100
All fresh input	Claude Opus 5.5 input at $4 per million	About $840
Worst case	All output on Claude Fable 5.1 at $50 per million	About $10,500

Delana estimate using Anthropic list prices as of September 2026 (Opus 5.5: $4 input, $20 output, $0.20 cached reads per million tokens; Fable 5.1: $10 input, $50 output). Real agent runs are mostly reading and caching, so the realistic figure sits at the low end. Actual cost depends on the model, caching and the input-to-output mix.

Even the unrealistic worst case is comparable to a couple of weeks of a senior specialist’s time, and a well-cached run costs a small fraction of a single day. That is why the pattern will spread well beyond biology.

The pattern your business can borrow

Strip away the DNA and the ART workflow is a repeatable loop: a person asks a sharp question, many agents search in parallel, the agents check and cut their own ideas, they write short evidence-backed reports, humans decide what deserves real-world effort, and what passes review sharpens the next brief.

Step	Owner	What happens
1. Ask	People	One sharp, high-level question
2. Search	AI agents	Hundreds of agents run in parallel
3. Self-check	AI agents	Reproduce known results and cut weak candidates
4. Report	AI agents	Short, readable write-ups with evidence attached
5. Test	People	Pick what is worth real-world effort
6. Refine	People	What passed review shapes the next brief

The agent discovery loop. People own the question and the decision; agents own the search, the first filter and the write-up.

Places where a small or mid-sized business could run this loop today:

Contracts and leases: scan every agreement for unusual renewal, liability or data-handling clauses, and surface the ten that deserve a lawyer’s attention.
Security logs and alerts: look for patterns that match no known rule, such as odd login timing or quiet data movement, and write up the few worth investigating. See our guide to AI-powered threats for SMBs.
Codebases: hunt for hardcoded credentials, dead integrations and risky dependencies across every repository, not just the ones someone remembers.
Customer feedback: read every ticket, review and call note to find complaints that cluster in ways no dashboard tracks.
Supplier and vendor risk: compare security questionnaires, certificates and public filings to flag the vendors whose answers do not add up.

The new bottleneck: human judgment

The most useful lesson from Anthropic’s lab is the one it did not headline. When search becomes cheap, review becomes the constraint. A swarm that produces 3,500 candidates is useless if nobody can evaluate them, and a swarm that produces 20 polished, confident reports is dangerous if nobody checks the evidence. Anthropic’s answer was to make the agents cut aggressively, require readable evidence in every report, and study its own reviewers’ choices. Businesses need the same discipline: decide in advance who reviews, how many reports they can handle, and what evidence a report must contain before it reaches them.

Five risks before you unleash a swarm

Runaway cost. Hundreds of agents burn tokens fast. Set hard budgets, caps and alerts per run. Our piece on AI cost overruns explains why inference bills are routinely under-forecast.
Confident but wrong reports. An agent convinced it has found something will write persuasively. Require citations to the underlying records and have a human spot-check them.
Data exposure. A swarm needs broad read access, which makes it a large target. Use least-privilege, read-only credentials, keep sensitive data in approved tools, and avoid the shadow AI trap of staff running searches in personal accounts.
Prompt injection. Agents that read emails, tickets or documents can be steered by hidden instructions inside them. The defenses are covered in AI Agent Security in 2026.
Dual use. The same capability that finds useful enzymes or fraud patterns could find harmful ones. Anthropic describes biology as dual-use and keeps experiments in human hands. Scope what agents may search for, log it, and keep action separate from discovery.

A six-step plan for your first discovery sprint

Pick one bounded question with a clear payoff, such as which of our 400 vendor contracts auto-renew with a price increase.
Validate on known answers first. Like Claude reproducing established results, have agents find issues you already know about before trusting them with unknowns. Run important searches more than once, since agent results vary from run to run.
Give read-only, least-privilege access to a copy of the data where possible, with every agent action logged.
Set the budget and the funnel up front: a token cap, a maximum number of reports, and the evidence each report must include.
Name the human reviewers and their capacity before the run starts, and make them the only path from finding to action.
Record what reviewers accept and reject, then feed those patterns into the next brief so the swarm learns your standards.

What to watch next

Whether independent labs confirm ART as a distinct system and identify what it does.
Peer review of Anthropic’s preprint and responses from the teams behind related systems.
Whether Anthropic moves toward agents operating lab equipment. Amodei said that may eventually be possible with safeguards, but that the company is not doing it today.
Similar agent-swarm results from other AI labs and from enterprises using the same approach on business data.
How AI providers price and cap very large multi-agent runs as they spread.

Frequently asked questions

What did Claude discover?

Anthropic says Claude agents identified a previously uncharacterized enzyme system in bacteriophages, called array-associated reverse transcriptases (ART). It combines an enzyme that copies RNA into DNA, a partner gene and an array of evenly spaced DNA repeats that resembles a CRISPR array. Its function is still unknown.

Is ART the next CRISPR?

Too early to say. It shares features with programmable systems that cut, copy or paste DNA, and early experiments show its array produces distinct short RNAs, but no one has shown what it does or whether it can be used as a tool.

Did the AI run the experiments?

No. Anthropic says its scientists wrote the initial prompt and performed all lab work. The agents searched the data, analyzed candidates and wrote reports for human review.

What is an AI agent swarm?

A large number of AI agents working in parallel on parts of the same problem, coordinated by a harness that collects and filters their output. In the ART search, about 950 agents worked for about 21 hours.

Can a small business use this approach?

Yes, at a smaller scale. The same loop of parallel search, self-filtering and human review works on contracts, logs, code and customer feedback. Start with a bounded question, read-only access, a hard budget and named reviewers.

Sources

Want to point an agent swarm at your contracts, logs or code without exposing your data or your budget? Delana can scope the question, set up least-privilege access and logging, and design the human review step. Talk to Delana.