Agentic Discovery: Inside the Harness That Found ART
- Read the ART campaign as an architecture, not a headline: five agent roles, a gated task chain, and a triage queue
- Understand why 98 of the campaign's 119 tasks were not in the brief — and why the discovery came out of a rejected hypothesis
- Parse a multi-agent token bill correctly: 189.5M of the 215.6M tokens were prompt-cache writes
- Explain the two findings the press coverage skipped — 0-for-10 reproduction, and tools making the model worse at seeing the data
- Copy the transferable parts of the harness for a non-biological corpus, and know which parts do not transfer
On September 23 2026 Anthropic published Claude discovers a novel enzyme system with CRISPR-like repeats. It hit the Hacker News front page the same evening and ran to roughly 655 points and 672 comments, plus wire coverage on every major outlet and a visible wobble in gene-editing stocks. The summary everyone repeated: ~950 agents, 21 hours, 210 million tokens, one new enzyme system.
Those numbers are real. They are also the least useful part of the story.
The accompanying preprint — Autonomous AI agents discover reverse transcriptases with tandem repeat arrays, by Yoon, Athukoralage, Ameisen, Kauderer-Abrams, Perry and Durrant — describes the harness, the accounting, the failure modes and a set of controls. It contains at least four things that change how you should build your own agent pipelines, and that essentially none of the coverage mentioned. This page is about those.
What the campaign actually was
A research brief — a short document stating the goal, not a workflow — was expanded by a launch agent into five stages, run as a chain with scripted completion checks between them. No task from a later stage could open until the previous stage closed.
| Stage | What it did |
|---|---|
| 1. Input assembly | Build reference sets and profile HMMs from scratch |
| 2. Database sweep | Survey reverse transcriptase (RT) loci across 1.9 billion metagenomic protein clusters |
| 3. RT classification | Sort ~200,000 recovered RT clusters into nine classes |
| 4. Neighborhood census | Sample ~11,000 loci, score 3,564 recurring neighbor protein families as candidate partners |
| 5. Deep dives | Investigate the families that passed the agents' own selection criteria |
The funnel that came out the other end: 3,564 candidate families → 16 passing the agents' criteria (+1 promoted later) → 19 written reports → 3 confirmed as genuinely new RT associations. Fourteen of the seventeen were thrown out as annotation artifacts, parts of already-described systems, or neighbors that just happened to live nearby. Separately, workers flagged three unusual RTs that turned out to be new lineages — one of which is ART.
That survival rate is the honest headline. A well-run autonomous campaign over a billion-scale corpus produced three defensible associations and one interesting anomaly. Budget accordingly.
The harness: five roles, one shared record
Every agent in the campaign was an instance of Claude Code configured with Claude Mythos 5, running in a sandbox with 60 CPU cores, 192 GiB of memory and no GPU — structure predictions were dispatched to a separate GPU queue (19 jobs, ESMFold or ColabFold). The harness allowed up to 58 concurrent sessions.
- One session turns the research brief into the five-stage chain, each stage closed by a scripted completion check. This is the only place a human intent enters the system — Anthropic states their involvement was the initial prompt and the lab work.
- Each task carries a task brief. The worker writes a plan, runs it, and submits a written summary together with the code and data files it produced. 414 of the 949 sessions were workers, accounting for 63.7 of the 76.9 agent-hours.
- A second agent reads the plan, the summary and the edited files, then accepts or returns the task for revision — 49 of 119 tasks were revised at least once. Critically, the supervisor can also open new tasks in response to what it just read.
- Follow-up tasks proposed by agents do not run automatically. They enter a triage queue from which the harness releases them or rejects them with a written reason. 10 of 119 tasks were rejected here.
- After each completed task a curator agent reads the brief and the summary and enters the findings into a shared knowledge base. Relevant entries are then injected into the prompts of later workers and supervisors. This is the mechanism that makes 949 sessions one campaign instead of 949 strangers.
- Reports pass an editor agent before entering the record. Everything — plans, results, reviews, scripts — is committed to a version-controlled store readable by every agent.
- There is no fixed step budget and no wall-clock stop. The run terminates when no dispatchable task remains — 21.5 hours later, in this case.
Session split across roles: 414 worker, 375 supervisor, 107 curator, 52 editor, plus the launch session — 949 total. Note that supervision costs almost as much as work. If you are sizing a review-heavy harness, a near-1:1 worker-to-supervisor session ratio is the empirical shape here, not a 10:1 one.
The agents had ordinary Claude Code tools (shell, files, web search) plus harness-added connectors to the protein database, to literature search and full-text retrieval, to the shared knowledge base, and to the GPU queue. They installed and ran standard bioinformatics software themselves — HMMER, MMseqs2, BLAST+, MAFFT, FastTree, Infernal, ViennaRNA and others. Across the retained logs: 7,578 shell commands, 696 database queries, 131 literature searches, 61 web requests.
One detail worth stealing: the sandbox held a library of about 140 skills — written method guides — of which workers consulted 15. Seven were survey-specific, and their titles read like a methodology checklist: counting independent occurrences of a gene family, scoring operonic arrangement against the window background, adjudicating candidate families, testing novelty claims, annotating uncharacterized families, proposing mechanisms together with the counterfactual evidence, and drawing gene-neighborhood figures. Those last two are the ones that keep an autonomous campaign honest.
98 of 119 tasks were not in the brief
The campaign's task record breaks down as: 5 tasks seeded from the brief's stages, 16 deep dives seeded from the promoted families, and 98 follow-up tasks generated during the run. Roughly four in five units of work were invented by the agents mid-campaign.
That is not overhead. It is where the result came from.
The chain, from the paper's transcript analysis:
- In task
t0010, a worker examined an RT family that had been selected because it sat next to a jumbo-phage RNA polymerase subunit. It rejected that association as spurious gene order — but queued a follow-up to look at the RT itself anyway, noting that a free-standing retron-like RT in a jumbo phage was a reportable secondary finding. - The supervisor reasoned that retron RTs work with a non-coding RNA encoded upstream, and wrote the next task brief: check the 5′ flanks for such an RNA.
- Those RTs had upstream regions with a median of 27 bp — far too short to encode one. Their closest relatives had a median of 940 bp. The worker pulled the relatives' upstream sequence directly into its context to look.
- In task
t0062, reading that raw sequence, it noticed by eye a tandem repeat array — a ~17 bp motif matchingCATGTG[T/A]TCGCATGT, repeating with ~110–180 bp spacers across a 2,900 nt flank of a 228,907 bp contig. A second locus showed the same architecture with a completely different repeat. - It then argued against itself — comparing the pattern with DGR variable repeats, CRISPR mini-arrays and known diversity-generating RT systems — before running a literature search designed to kill the novelty claim rather than confirm it.
Neither the research brief nor the task brief mentioned repeats or arrays. No repeat-finding tool was run in that session. The finding exists because a rejected hypothesis was allowed to spawn a follow-up, and because the harness let a worker chase something it was not asked about.
Design implication: if your agent pipeline only executes the DAG you drew, you have built a slower version of a conventional pipeline. The parts to add are the ones that cost money and look like waste — a channel for agent-proposed work, and a gate that reviews those proposals in writing instead of auto-running or auto-dropping them.
Read the token bill correctly
The widely-quoted figure is "210 million tokens". The paper's accounting is more specific, and the shape of it matters:
| Component | Tokens |
|---|---|
| Uncached input | 11.3M |
| Output | 14.9M |
| Written to prompt cache | 189.5M |
| Reported total | 215.6M |
| Read from prompt cache | excluded from the total |
So 88% of the headline number is cache writes, and the largest real quantity in the system — cache reads, which is how 949 sessions were fed a shared knowledge base — is not in it at all. Output tokens, the thing people instinctively estimate, are 7%.
Two practical consequences. First, if you benchmark your own harness against "210M tokens", make sure you are counting the same four things; most dashboards report a different mix. Second, in a many-session harness with a common memory, cache economics dominate cost — a cache write is billed at a premium over base input and a cache read at a steep discount, so the ratio between the two is the single biggest lever on the bill. See Prompt Caching Economics for how that arithmetic works. The paper does not state a dollar cost, and neither will we.
Wall-clock was 21.5 hours against 76.9 agent-hours — an average concurrency of about 3.6 despite a ceiling of 58. Long-running agent campaigns are far more serial than the concurrency limit suggests, because stage gating and review loops force waits.
The finding did not reproduce
This is the result that should be quoted more than any other.
The authors relaunched the identical campaign, same harness and same brief, ten more times, producing 3,084 task records and 5,632 session transcripts. Nearly every replicate that completed the census sampled ART loci. In two of them, workers opened follow-ups on the lineage.
None of them read the DNA upstream. The array was missed in every rerun. The paper attributes this to the breadth of the RT search space and the non-deterministic behavior of the harness.
Take that seriously in both directions:
- Against over-claiming. "AI discovered X" describes a single sample from a stochastic process, not a capability you can invoke. An 11-run campaign found it once.
- For designing pipelines. The unit of an agentic search is not the run, it is the population of runs. If a finding is worth anything, replicate the campaign, dedupe across replicates, and treat single-run reports as leads. This also gives you a free negative control: anything every run produces is probably a property of your pipeline, not of the data.
- For evaluation. A harness that surfaces a real anomaly 1 time in 11 is still enormously valuable if the cost of a run is a few hundred dollars and the payoff is a new biological system. Low reproducibility is not the same as low expected value — but you have to price the run, not the result.
Giving the model tools made it blind
Because the campaign itself was not reproducible, the authors built a fixed-input benchmark: give one model the ART system at one of five information levels, ask it to characterize the locus and predict its function, and have a judge model (Mythos 5) score the report against ten curated claims about the array, the RT and the partner.
The five levels escalate from two RT protein sequences, to two loci as raw text, to all 96 loci as files with tools, to files plus predicted structures, to files plus literature and web access.
The expected result is that more information and more tooling monotonically helps. It did, for the partner gene. For the array — the actual discovery — it inverted:
- With the loci in context as text, the most capable models described the array correctly in at least 90% of attempts.
- With the same data as files, plus tools, recognition fell to as low as 32% (Opus 5, level 4).
- Inspecting transcripts explained it: in the file-based runs, 39% of attempts never read a contiguous stretch of 200 nt or more — never seeing more than about one repeat unit. They queried, grepped, summarized and reasoned about the sequence without ever looking at it.
- When a run did read ≥200 nt contiguously, recognition was 16 to 32 percentage points higher. Pooled across the four strongest models, recognition rose from 29% to 76% as more DNA entered context, and reached 96% for Mythos 5.
The model panel itself is worth noting: the paper tested seven Claude models and found a clear capability gap between the four strongest (Opus 5.5, Mythos 5.1, Mythos 5, Opus 5) and the rest (Opus 4.8, Opus 4.6, Sonnet 5) — evidence that the original campaign's success depended on the frontier model it ran on.
The transferable lesson has nothing to do with biology. Whenever the signal lives in the shape of raw data — a repeating pattern in a log, a regular offset in a binary, an odd rhythm in a time series, duplicated stanzas in a config — tool-mediated access lets the model do the reasonable thing and summarize instead of look. The tool is an abstraction, and abstractions are exactly what hide anomalies. This is the opposite of the usual context-engineering advice, and it is a real exception to it:
Force direct observation when you're hunting anomalies
When investigating this record, do not rely only on summaries, counts, greps, or tool output. Read at least 2,000 contiguous raw bytes/characters of the region itself into your context and describe, literally, what you see — including anything structural you notice that I did not ask about. Then state: (a) what you expected this region to look like, (b) what it actually looks like, and (c) any difference between the two that you cannot explain. Treat (c) as the finding, not as noise.
Pair that with the standard advice in Context Engineering: curate aggressively for execution tasks, but pay the tokens to look directly at primary data for discovery tasks. They are different jobs.
Why it noticed: a detector that fires on repeats
The paper's most unusual section is an interpretability analysis run on the actual discovery transcript. The authors passed the worker's session back through Mythos 5 and decomposed the model's activity at one layer into interpretable signals, looking for something that responds to repeated DNA.
They found two. Both were silent before the array began. Repeat-signal 2 strengthened by the third repeat copy and stayed steady across the rest, with activity confined almost entirely to the repeat and the nucleotide immediately after it. Repeat-signal 1 peaked at the third copy and then faded. The control is the convincing part: shuffling the nucleotides within each copy — preserving composition, destroying the repetition — silenced repeat-signal 2 on all 14 copies and repeat-signal 1 on 12 of 14. Neither signal is specific to DNA.
For comparison they ran two dedicated genomic language models, Evo 2 and gLM2, over the same locus. Both flagged the repeats confidently, and the bidirectional model (gLM2) caught every copy including the first, while the causal one (Evo 2) predicted the first two copies poorly and improved along the array — exactly the behavior you would expect from a model that can only see backwards.
In the transcript, the signal firing is immediately followed by natural-language reasoning that names the pattern a tandem repeat array. A general-purpose assistant carried an internal detector for a structural property of raw sequence, and that detector's activation surfaced as a verbalized observation. That is a concrete, mechanistic answer to "how did it notice?" — and a reason to believe the tool-blindness result rather than dismiss it: a detector cannot fire on bytes that never entered the context.
The judge did not rank it first
The 19 reports were ranked by a tournament of pairwise comparisons — every ordered pair compared once, 342 games, with a Mythos 5 judge reading a fixed rubric and two anonymized reports and scoring each from 1 to 5 on impact (weight 0.35), novelty (0.30), soundness (0.25) and actionability (0.10). A report scoring 2 or less on soundness lost automatically. Ratings were fitted with a Bradley–Terry model.
The ART report finished third, with 32 wins in 36 games.
Three things to take from that. Anonymized round-robin with an auto-loss on soundness is a much better LLM-judge design than scoring reports independently on a 1–10 scale — pairwise comparison is more stable, and the soundness floor kills confident nonsense before it can win on novelty. But the best result in the batch still was not ranked first, so treat an LLM-judge ranking as a triage order, not a verdict, and keep your top-k wide enough that a third-place item still reaches a human. And note that the whole apparatus is downstream of humans anyway: every candidate that survived went to human scientists, and the wet-lab work was entirely human. See The Verification Bottleneck for the general shape of this constraint.
What is actually known about ART
Enough hedging is warranted here that it is worth separating cleanly.
Supported by the preprint:
- ART — array-associated reverse transcriptases — is a family of RTs largely confined to jumbo-phage genomes, sitting in a clade next to retrons.
- Each locus has three parts: an array of ~200 nt repeat units, the RT, and a dedicated partner gene. Three structural types of partner protein were observed (59, 7 and 3 lineages).
- The repeat layout resembles a CRISPR array, but unlike CRISPR spacers, ART units are conserved between related phages.
- In an RNA-seq time course of Staphylococcus phage SA1 infection (public BioProject PRJNA836150), the array is transcribed abundantly and appears as discrete units — transcribed in large excess over the enzyme itself. That pattern suggests one enzyme supplied with a repertoire of distinct RNAs.
- The system was hiding in plain sight: for phage MarsHill (GenBank
MW248466.1), the original genome report identified the RT and proposed a 5′ non-coding RNA but described neither the repeats nor the partner gene.
Not shown:
- That the RT is enzymatically active.
- That the array RNAs are its substrates.
- That the RT and its partner interact at all.
- What the system does for the phage. This is open, and the authors say so.
Anything framing ART as "a new gene-editing tool" is running several years ahead of the evidence. What has been demonstrated is the discovery process, not the biology.
Copying the shape for a non-biological corpus
The domain-specific half of this — metagenomic databases, HMM profiles, structure prediction — does not transfer. The harness does. If you are running agents over a large corpus you do not fully understand (logs, a legacy monorepo, contracts, telemetry, a package registry), this is the pattern:
- State the goal and the criteria for something being interesting. Let a launch agent expand it into stages. If you specify the steps, you get a pipeline — and a pipeline can only find the novelty you defined in advance.
- No task from stage N+1 opens until stage N passes a deterministic completion check. This is what stops a 900-session run from compounding an early mistake through every later stage.
- Different sessions, different prompts. The supervisor reviews the plan and the artifacts, not just the summary, and can send work back — about 40% of tasks here were revised at least once.
- A triage queue that releases or rejects with a written reason. This is where the actual finding came from, and where an unbounded run would otherwise wander off.
- A curator role that writes each task's findings into a shared store, with relevant entries injected into later agents' prompts. Storage nobody reads is not memory.
- Explicitly require contiguous reads of primary data in the task brief for any anomaly-hunting stage. Tools invite summarization; the 90%-to-32% gap is what that costs.
- Fixed rubric, weighted dimensions, automatic loss for low soundness, Bradley–Terry fit. Then hand the top-k — not the top-1 — to a human.
- Same brief, several runs. Findings that appear once are leads; findings that appear every time are usually artifacts of your harness. Here, the real one appeared once in eleven.
Test yourself
Check yourself
0/5Sources & further reading
- Anthropic — Claude discovers a novel enzyme system with CRISPR-like repeats (Sept 23 2026): https://www.anthropic.com/news/claude-discovers-novel-enzyme-system
- Preprint — Yoon, Athukoralage, Ameisen, Kauderer-Abrams, Perry, Durrant, Autonomous AI agents discover reverse transcriptases with tandem repeat arrays (Anthropic, Sept 2026). All harness, accounting, replicate, benchmark and interpretability figures on this page are from its Results and Methods: https://www-cdn.anthropic.com/22573675ada52a8ca8a97a1a4b4326b2f208a071.pdf
- Hacker News discussion (655 points, 672 comments, Sept 23 2026): https://news.ycombinator.com/item?id=49820134
- Anthropic — Claude Science, an AI workbench for scientists (June 30 2026), the platform half of the toolchain: https://www.anthropic.com/news/claude-science-ai-workbench
- Anthropic Engineering — Effective harnesses for long-running agents: https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents
- Arc Institute — Evo 2, the genomic language model used as a comparison for repeat detection: https://github.com/ArcInstitute/evo2
- RNA-seq data behind the expression result: NCBI BioProject PRJNA836150 — https://www.ncbi.nlm.nih.gov/bioproject/PRJNA836150 — and the MarsHill phage record
MW248466.1: https://www.ncbi.nlm.nih.gov/nuccore/MW248466.1
Next
- Harnesses for Long-Running Agents — the general pattern this campaign is one instance of
- How Agent Memory Actually Works — what the curator role is doing, in the wider taxonomy
- Claude Science — the workbench the campaign ran on
- Prompt Caching Economics — how to price a 189.5M-cache-write run