950 Agents, 21 Hours, 200,000 DNA Sequences

After Anthropic established its life sciences research group this year, it published its first results. They deployed 950 Claude Agents into a massive DNA database to identify retrotransposon systems that hadn’t been clearly characterized before. Retrotransposases do something specific: they copy RNA information back into DNA.

Over 21 hours, these agents consumed 210 million tokens, collected over 200,000 retrotransposons, filtered down to 3,500 candidate systems, and converged on 20 priority targets for deeper analysis.

The turning point came during a routine check. One agent, reviewing DNA adjacent to a candidate, noticed a series of neatly arranged repeats and stopped. The arrangement looked like CRISPR. Phage retrotransposases had long been studied, but no one had noticed the unknown-gene neighbor and the long stretch of evenly spaced DNA repeats. Claude treated the three elements as one integrated system. Anthropic named it ART—Array-Related Retrotransposase. Follow-up experiments revealed that ART arrays similarly express a set of short RNAs, and a potentially programmable mechanism may be embedded within them.

Note the boundary. Claude’s job was database search, candidate filtering, and analysis. All wet-lab work was done by human scientists. Whether ART can cut, copy, and paste DNA remains unanswered.

A Pharma Company with 37,000 Virtual Employees

A week before Claude’s ART discovery, Stanford University School of Medicine executed something even larger. They built a virtual biotech company structured like a real pharma organization: 37,000 AI employees, a chief science officer overseeing different departments, each tasked with finding drug targets, analyzing clinical data, designing drugs, and planning trials. No offices. No salaries. No lunch breaks.

Researchers assigned approximately 50,000 clinical trials to the agents, one per agent, and the system aggregated the results. In less than a week, they identified a set of features correlated with drug success rates: targets concentrated in specific cell types, and gene activity that behaves more like an on-off switch than a dial capable of gradual adjustment. Drugs with these characteristics had higher probabilities of progressing from Phase I to Phase II and ultimately reaching the market, with fewer side effects.

Next, they had the agents design a treatment strategy targeting B7-H3 in lung cancer. The agents proposed antibody-drug conjugates—molecules that home in on B7-H3-positive cells and deliver chemotherapy directly to their vicinity. A few months later, a real pharmaceutical company independently proposed the same therapeutic approach, and the drug subsequently received FDA Breakthrough Therapy designation.

The third line of work has already entered human trials. Insilico Medicine’s Rentosertib for idiopathic pulmonary fibrosis saw AI-assisted target discovery (TNIK) and candidate molecule generation and screening. New research published this September reanalyzed blood samples from the Phase II trial: 42 participants, 12 weeks of complete data, six protein-based aging clocks. All six clocks showed changes in the same direction—patients’ predicted biological ages decreased after treatment. The researchers themselves acknowledged they couldn’t fully disentangle whether the changes stemmed from disease improvement or from the drug genuinely intervening in aging.

Two Cases, One Pattern

Laid side by side, the three cases share an identical structure:

Machines produce “possibilities.” Humans make “judgments on possibilities.” The original quote nails it: generating hypotheses is speeding up, but validating hypotheses still requires waiting for cell growth, animal studies, ethics review, and human trials. AI can produce thousands of seemingly plausible answers at once. The bottleneck becomes whether we have enough experiments, time, and patience to confirm which ones are real.

This is the sole criterion for individual developers choosing their domain.

Three Filters for Choosing a Domain

I wanted to build a workflow of thousands of agents that continuously surfaces things humans need a great deal of time to find. With a CS undergraduate and master’s background, I’m filtering domains through three practical criteria.

Filter 1: Validation cost. When an agent produces 1,000 candidates, how much does it cost and how long does it take to confirm each one? Database queries are nearly free. Reproducing a vulnerability requires running a PoC. Wet-lab experiments demand cell cultures. This is the first filter: domains with expensive validation get deprioritized.

Filter 2: Search space. The total object count must be large enough that humans can’t exhaust it. 200,000 retrotransposons, 50,000 clinical trials—at that scale, agents get to play. In domains with small search spaces, humans are sufficient on their own.

Filter 3: Data accessibility. Can all target objects be obtained programmatically in batch? Public databases, open-source repos, and published court opinions all qualify. Anything requiring authorization, web scraping, or purchase adds cost upfront.

Applying these three filters, here are the domains accessible to a CS background:

Four Lines Already Running

Vulnerability Bounties: The Most Expensive Validation

This line has been running for months: a multi-agent pipeline selects candidate open-source repos, greps for dangerous sinks, runs local minimum reproductions, then writes and submits reports. GitHub Security Lab and huntr bounties are the monetization outlets, with Google OSS VRP as a supplement.

The advantage: all data is public. A single GitHub repository has everything you need—every line of code. The bottleneck is equally clear: a suspected vulnerability only becomes bounty-eligible when reproduced locally. AI can assist, but the final call rests with the submitter. Anthropic handing its 20 priority targets to human labs, and my PoC needing to pass before submission—these are the same move.

Scaling to thousands of agents has a clear direction: the broad-screening layer can be parallelized indefinitely. A single pipeline can simultaneously monitor dependency changes and diffs across hundreds of repos, turning “newly introduced dangerous calls” into a persistent event feed. The convergence layer stays at dozens. The validation layer stays manual. The broad-screen-to-convergence ratio, following Anthropic’s 200,000-to-20 benchmark, is roughly one ten-thousandth.

OSINT Consolidation: The Cheapest Validation

Background checks and competitive monitoring I’ve built for enterprises follow this path. Multiple agents pull data from business registries, judicial records, bidding announcements, recruitment platforms, and news, then synthesize everything into evidence cards—with every fact citation-tracked. The adjudication layer uses multi-model adversarial validation plus an independent judge: a fact only counts after passing two independent checks.

Validation cost here is near zero, because the validation method is cross-referencing: three independent sources agreeing yields sufficient confidence. This mirrors the Stanford virtual pharma model—departmentally organized agents producing output for humans to make final calls.

The weakness is information-asymmetry monetization: a conclusion’s value depends on how well the question is asked. “Is this company worth partnering with?” is ten times more valuable than “How is this company?” Productization needs to lean into specific decision scenarios—investment due diligence, vendor onboarding, partner risk control.

Policy Matching: The High-Information-Asymmetry Mid-Range

Government subsidies, park residency benefits, and talent certification policies—marketing articles screaming “up to ¥1 million” and “up to 3 years free”—need to be decomposed into actual eligibility criteria. The verification workflow I built splits every promise into three blocks: qualification criteria, fulfillment pathways, and application costs, then cross-references each line against the official source text.

National policy databases are scattered across government websites at every level—large enough in scale, fully public. The agent’s job is full-coverage scanning and structuring, turning “who qualifies for what policy” into a matching engine. Validation relies on official documents and phone calls—medium cost.

The information asymmetry here is large enough to warrant the effort. The very reason it’s worth pursuing is that the search space is filled with unorganized material. It’s a strong fit for individual developers with low cold-start costs.

Football Data: The Slowest Compound Interest

Match video converted to structured event JSON—events only, no video storage, sidestepping copyright. Positioning this as a 5-to-10-year refinement project, anchored in Spain’s Position Play methodology, differentiated by exclusive amateur-league data assets.

The agent’s role here differs from the first three lines. This isn’t search—it’s annotation and mining. Videos first become events; once the event database reaches sufficient scale, agents can start answering questions like “what’s the success rate of this kind of possession-based play in amateur matches”—questions nobody else can answer.

Validation relies on humans watching replays—medium cost. The value here is compound interest: the longer you invest, the thicker the data asset becomes, and anyone starting from zero must re-record everything. The imagination space for a B2C data supply chain lives here. There’s no rushing this.

Two More to Watch but Not Yet Entered

Patent & literature blind spots. Scan the full corpus of patents and papers using the Anthropic-ART approach, surfacing overlooked technology combinations. Validation is a search confirming “has this combination actually been done before?"—nearly free. The门槛 here is monetization: the discovery itself isn’t valuable. What’s valuable is discovery plus a path to implementation. Worth opening as an experimental line once the existing four have pipeline capacity to spare.

Drug clinical data reanalysis. ClinicalTrials.gov is fully public; Stanford has already demonstrated the approach. Breaking into this with a CS background requires补 in biology domain knowledge—the补课 cost is high. My judgment: leave this to others. Tracking methodological advances from the sidelines is more cost-effective than jumping in.

The Sort Order

Cash flow comes from vulnerability bounties. Validation is expensive, but returns are the most direct—and it’s already producing. Product and company形态 come from OSINT consolidation and policy matching. Validation is cheap, information asymmetry is large, and both can become B2B services. Football data is a ten-year asset reserve—keep investing, don’t expect near-term returns.

Thousands of agents aren’t the goal. They’re the means. In Anthropic’s case, 200,000 DNA sequences were whittled down to 20 for humans. When the ratio is right, scale becomes meaningful.

One lingering fact: that virtual pharma company with 37,000 employees organized 50,000 clinical trials in less than a week. The B7-H3 proposal, from agent generation to independent validation by a real pharma company, spanned several months. The gap is right there—the time difference between machine computation and human validation is three orders of magnitude. What thousands of agents are worth depends on how much you can compress those three months.