ReMyndclaude enzyme discovery: did agents find a new crispr-like system?
as of september 2026, the claude enzyme discovery story is this. anthropic's agents found a previously undescribed phage system called art, not a clinic-ready crispr-cas editor. the september 23, 2026 anthropic post says the agents spotted a reverse transcriptase beside a crispr-like dna repeat array. humans still do not know what art does. humans still ran every wet-lab experiment.
that is a real agentic result, and a narrower one than the headlines suggest.
what did anthropic announce on september 23, 2026?
anthropic opened an anthropic life sciences lab program and shared the first early result from it.
the group formed in spring 2026 to ask whether general models can help find uncharacterized protein families in dna databases, then hand candidates to scientists for lab tests. the first public case is the claude art enzyme: array-associated reverse transcriptases, found mainly in bacteriophages.
| claim | what anthropic says | status |
|---|---|---|
| new crispr-cas enzyme | no. art has crispr-like repeats, not cas genes | clarified |
| novel enzyme system | yes. array + partner + rt together were not described this way | claimed, preprint |
| function known | no. primary function is still under study | open |
| lab work by claude | no. humans did bsl-1 and bsl-2 lab work | stated |
| autonomous database search | yes. agents ran the genome-mining campaign | claimed, preprint |
the underlying reverse transcriptase in a jumbo phage had shown up in earlier studies. anthropic's judgement is that claude appears first to have noticed the defining extras: the non-coding repeat array and an accessory partner protein. the next web reported the same caveats on september 23, 2026.
how did the claude agents find art?
the campaign was multi-agent mythos 5 genome mining, not a single chat that spat out a paper.
anthropic gave a research brief: search about 1.9 billion protein clusters for interesting reverse transcriptase systems. worker and supervisor agents in claude code, running claude mythos 5, planned searches, reviewed each other, and opened follow-ups when something looked odd.
numbers from the preprint pdf:
- about 200,000 rt clusters recovered, then classified into 9 classes
- about 11,000 loci sampled and 3,564 recurring neighbor families scored as partners
- 17 families promoted for deep dives, then 19 written reports ranked for humans
- 119 tasks, 949 agent sessions, 215.6 million tokens, 21.5 hours wall clock with 0 human interventions during the run
of the 119 tasks, 98 were agent-proposed follow-ups, or about 82% of the queue. art arrived on a side path. an agent reading raw upstream dna wrote that it could see a tandem repeat array by eye, counted repeats and spacers, compared known systems, searched the literature, and filed a report.
feng zhang of mit and the broad institute, after reviewing the preprint, called it an exciting example of how ai agents can contribute to biological discovery. that quote is praise for the process, not a clinical claim about the enzyme.
what is an array-associated reverse transcriptase, and what is still unknown?
art is a structural description plus a hypothesis, not a shipped gene editor.
anthropic's follow-up analysis found 95 distinct art-family rt clusters at 90% identity, with a detectable repeat array upstream in 28 of them (about 29% of the family). arrays span roughly 0.3 to 4.1 kb and hold 3 to 21 short repeats at near-constant spacing. spacers run about 120 to 220 nt and are unrelated within an array. 0 cas genes sit beside these loci.
early experiments matter, and they stop short of a mechanism:
- the array is expressed as distinct short rnas
- in published staphylococcus phage sa1 infection rna-seq (microorganisms 2022), array-derived rnas reached as much as 8% of phage rnas at 15 minutes post infection
- anthropic has not shown the rt is active, or that those rnas are its substrates
- whether the rt and partner interact, and what the system does for the phage, remain unknown
so if you searched "did claude invent a new crispr", the accurate answer is no. if you searched for ai agents biology discovery that notices an anomaly humans missed in a huge sequence dump, anthropic's evidence says yes in this case, with the usual preprint caveats.
did the discovery replicate cleanly?
replication is the uncomfortable part of the story, and anthropic published it.
the team re-ran the same campaign 10 more times. almost every completed census sampled art loci, and workers investigated the lineage in 2 runs, but 0 of the reruns read the upstream dna, and all 10 missed the array. the authors blame search-space size and non-deterministic harness behavior.
fixed-input benchmarks tell a clearer story. when the four most capable models tested (opus 5.5, mythos 5.1, mythos 5, opus 5) got the loci as text in context, they described the array in at least 90% of attempts.
with files and tools, recognition fell as low as 32% for opus 5 in one setting. often the model never pulled a long enough dna stretch into context. when at least 200 nt of contiguous locus dna was read, array recognition rose by 16 to 32 percentage points.
that is an ai agent lesson as much as a biology lesson: the model has to see the raw evidence, not only a summary of it. the same pattern shows up when people try to give coding agents memory of their own repos in giving an ai coding agent memory of your actual work.
what does this mean for research agents in 2026?
this result strengthens long-horizon agents on primary data, and it does not retire human taste or the bench.
use the claude enzyme discovery as a checklist if you run research agents yourself:
- write a brief, not a wish. anthropic started with a narrow mission: novel rt systems by partner associations.
- let agents open follow-ups. 98 of 119 tasks were agent-proposed extensions (about 82%).
- keep a shared record. plans, results and reviews went to a common memory the harness could reread.
- rank outputs for humans. 19 reports went through a tournament before scientists picked art.
- put humans on irreversible steps. anthropic states all lab work is human, at bsl-1 and bsl-2, with no human-infecting pathogens.
for ai for researchers, the gap is still personal context. anthropic's agents searched a shared metagenomic store. your own unread pdfs, rejected hypotheses and wet-lab notes are a different archive. how to feed that archive into an agent without pasting your life into a chat is the topic of how to give an ai agent context about your own work and what is context engineering.
verdict: treat art as a strong example of agentic anomaly detection in sequence data as of september 2026. do not treat it as a finished biotech product, and do not skip the preprint's open questions.
keep the lab notes on your side
a discovery campaign this long produces a trail: prompts, rejected candidates, rna plots, and the moment someone decided a flank was worth reading. that trail is as valuable as the final report. it usually lives in terminals, browsers and notebooks rather than in a reference manager.
remynd is a mac app built for that trail. it captures the focused window, runs ocr locally with apple vision, and keeps a searchable index on your mac. you can exclude apps and sites, and recordings are kept for 30 days by default. call transcription runs on device with mlx.
the agent window runs your own claude code or codex cli against that history with read-only access. settings accept a custom api endpoint compatible with the openai responses api. sign-in and cloud agents still use the network, so the honest claim is that the archive lives on your mac. the security page is at /security/, and the privacy details are in private ai on your mac.
download remynd for mac and keep the next research trail searchable.