Claude discovers a novel enzyme system with CRISPR-like repeats

TL;DR Highlight

950 agents over 21 hours surfaced a CRISPR-like novel enzyme system candidate

Who Should Read

Backend and agent developers designing pipelines that run many parallel agents to find anomalous patterns in large datasets

Core Mechanics

High-level prompt, parallel agents
- The prompt was to search a DNA sequence database for new reverse transcriptases (RTs, enzymes that copy RNA into DNA).
- Scientists say they handled only the initial prompt and lab work.
21 hours, ~950 agents, 210 million tokens
- One agent spotted a repeating DNA pattern next to an odd-looking RT gene.
- Further analysis and lab tests showed it marks a previously uncharacterized enzyme system.
The candidate funnel narrowed step by step
- Agents gathered over 200,000 RTs and picked 3,500 new candidate systems.
- They narrowed those to the 20 most compelling and wrote human-readable reports.
The agent followed a human-like verification routine
- It counted repeats, measured spacing, compared the layout with known RT systems, and searched the literature.
- Once convinced, it filed a report for human review.
The new system, ART, has three parts

Community Reactions

Commenters said the RT itself is already known
- Some said this is a new genomic arrangement around a known retron-like RT, not a brand-new system.
- Others said anyone with money could have gotten the same result.
Criticism of the "Claude found it" framing
- Many commenters said human scientists' contributions are hidden.
- One noted the linked technical report does credit the humans involved.
Commenters said harness and human judgment are what matter
- They argued data access and the ability to judge results are key, and that other models with a custom harness could do this.
- They pointed to the authors admitting to using their own harness.
Concern about filtering bias
- Commenters worried that narrowing 200k to 3,500 to 20 to 1 could discard genuinely novel candidates.
- They noted agents may favor things resembling known systems.
Lack of rigor and unshared prompts

How to Apply

Have agents write candidate reports, then rebut them
- Today, pick one pile of logs or documents and have an agent pull anomaly candidates.
- For each, write a short report with a proposed cause and evidence, then have a separate session critique it to prune candidates.
Split agents in parallel with a staged funnel
- Shard large data across agents and shrink the candidate count at each stage.
- Log candidate counts per stage to see where candidates get dropped.
Feed accepted and rejected candidates back into instructions
- Record which candidates you accept or reject, with reasons.
- Put those criteria into the next run's instructions.
Spot-check discarded candidates
- Sample some candidates dropped at early stages and review them yourself.
- Check whether novel ones were dropped just for looking unlike known patterns.

Try It

copy & run

# Paste into Claude or ChatGPT (simplified version of the paper's method: candidates → reports → critical review)

You are a data exploration agent. Find unexplained anomalous patterns in the data below.

[Data]
{{logs/documents/data}}

[Steps]
1. First reproduce known patterns from the data to verify your method works.
2. Pick up to 5 candidates that fit no known pattern.
3. Write a short report per candidate: proposed cause, supporting evidence, how it could be falsified.
4. Then act as a critical reviewer: rebut each report's evidence and keep only the candidates that survive.
5. If no candidate survives, say so.

Next to try: also output the rejected candidates and the reasons so a human can check them.

Terminology

Reverse transcriptase (RT)An enzyme that reads RNA and copies it into DNA. Think of a converter that rewrites a document into another format.

CRISPRA repeat-sequence structure found in bacterial immune systems that is now the basis of gene editing tools. It stores a bank of different RNA sequences, which lets it target chosen locations.

BacteriophageA virus that infects bacteria.

Genome miningSearching sequence databases for uncharacterized genes and picking out unusual ones. It's like grepping a codebase for odd functions.

HarnessA runtime framework that coordinates and manages many model sessions running in parallel.