10 min read - Anthropic’s enzyme search: what one result in eleven means for AI agents

Anthropic announced on 23 September 2026 that a campaign of Claude agents had identified a previously undescribed enzyme system in bacteriophages. The company calls it array-associated reverse transcriptase, or ART. Its announcement says the system’s function is still unknown, and all laboratory work was performed by human scientists.

Those qualifications are more useful to an AI buyer than the word “discovery.” They separate four different achievements: searching a large body of data, noticing an anomaly, deciding that it merits attention, and proving what it does. Anthropic’s agents did the first three. Human researchers handled the physical experiments and interpretation that turned a candidate into a scientific result.

The technical report says the researchers ran the same campaign ten more times, and every rerun missed the repeat array that defined ART. A standout result occurred, but the complete workflow did not reproduce it.

For a company considering agents for engineering, research, due diligence or operations, this is a good case study in how to assess a promising pilot. A demonstration can establish possibility. An investment decision needs evidence about yield, repeatability, verification effort and cost.

What the campaign established

The 40-page technical report describes a large search rather than a single chat. A research brief became a queue of tasks. Worker agents proposed and ran analyses; supervisor agents reviewed plans and results; curators updated shared knowledge; editor agents prepared reports for scientists.

The original campaign covered 119 tasks and 949 agent sessions. It used 215.6 million tokens and ran for 21.5 hours of wall-clock time. The agents recovered roughly 200,000 reverse-transcriptase clusters, examined candidate associations and filed 19 reports. One follow-up path led an agent to read raw DNA beside an unusual reverse transcriptase and notice a tandem-repeat pattern.

The researchers then did work the agents could not close. They reviewed the reports, checked the literature, named the family and ran sequencing experiments. Their evidence shows that ART arrays are transcribed into distinct short RNAs. It does not yet establish ART’s biological function or prove that the system is a programmable gene-editing tool. The resemblance to CRISPR concerns the repeat-array architecture, not an established equivalent capability.

Business summaries should preserve that boundary. “The agent surfaced a candidate that survived expert and experimental review” is accurate. “The agent produced a new biotechnology” is not.

The ten misses change the interpretation

In the ten reruns, nearly every campaign sampled ART loci and two investigated the lineage further. None read the upstream DNA in the way that exposed the repeat array. The report attributes the misses to a broad search space and non-deterministic agent behaviour.

This was not a simple model failure. The necessary data existed, relevant loci entered the campaign, and some agents came close. The workflow still failed to place the decisive raw sequence in the model’s context. Tool access and task completion therefore overstated the system’s effective coverage.

Anthropic’s fixed-input tests make that distinction clearer. When the most capable models received the relevant loci directly in context, they recognized the array in at least 90% of attempts. In tool-using setups, recognition fell as low as 32% for one tested configuration. Across the file-based attempts, 39% never read a contiguous stretch of at least 200 nucleotides, so they did not see more than about one repeat unit.

Similar blind spots appear in business workflows. A contract-analysis agent may access every document yet skip an annex. A coding agent may have repository access but never open the configuration that changes the answer. A support agent may query the ticket system while missing the attachment that contains the failure trace. Permission to inspect a source is not evidence that the workflow inspected it.

Measure the funnel, not the best output

An agent pilot should produce a funnel that a decision-maker can inspect. Start with the number of items available, then record what the system retrieved, what it actually read, which candidates it promoted, how many experts rejected, and what survived external verification.

For a discovery workflow, a compact evidence record could include:

Do not turn one success in eleven campaigns into an estimated success probability. The runs may not be independent, the search process adapts, and this sample is too narrow for a procurement forecast. It does support a simpler conclusion: the observed end-to-end result was not repeatable under the reported rerun conditions.

The same caution applies to pass-at-many metrics. If ten attempts give a team one useful answer, that can still have value for exploration. It is a poor basis for an unattended workflow that must return the right answer every time. Buyers should define whether their use case rewards occasional discovery, dependable completion, or both before comparing systems.

Keep verification outside the agent’s scoring loop

The ART work did not accept an agent report as biological truth. Human scientists used laboratory evidence to test the candidate. That separation matters in less scientific settings too.

If the agent proposes a security finding, reproduce it in a contained environment. If it extracts a contractual obligation, ask an authorised reviewer to compare the result with the signed document and applicable context. If it recommends a software change, run deterministic tests and review the affected boundary. The final check should measure the external object, not ask the same agent to grade its own answer.

NIST’s AI Risk Management Framework Core calls for documented, repeatable testing, evaluation, verification and validation. It also recommends recording knowledge limits, defining human oversight and involving assessors who did not build the system. Those practices fit the lesson here: the candidate-generation engine and the acceptance decision need separate evidence.

Independent review does not mean that every result requires a committee. Match the check to the cost of being wrong. A low-impact research lead may need a quick specialist screen; a decision affecting money, customers, safety or legal duties needs an authorised owner and a stronger test.

Instrument what the agent actually sees

The gap between access and observation is an engineering problem. Log retrievals, opened files, query parameters, tool errors and the source references used in the final report. Keep enough detail to explain why one run found an item and another missed it, while respecting the data rules for the system being tested.

Design a small coverage test before running the full pilot. Place harmless, known features across ordinary documents, awkward formats, linked attachments and deep directory paths. Then measure which features the agent retrieves and reads. A correct answer on the easy document does not compensate for silent misses elsewhere.

The evaluation-driven development guide explains how to turn examples into versioned tests. The agent observability guide covers the traces needed to investigate a miss. Together, those practices let a team compare changes to the model, prompt, tools and orchestration instead of treating the model name as the whole system.

Price the verification burden

The report publishes token use and elapsed time, but it does not provide a monetary cost for the campaign. It also does not reduce the subsequent work to a single number. Scientists had to review candidates, design experiments, run laboratory procedures and interpret the result.

That missing cost is common in agent business cases. Model usage is easy to meter, while rejected leads, specialist review, reruns and incident handling sit in other budgets. Calculate the cost per verified useful result, not the cost per generated report.

A useful pilot comparison has two baselines. Compare the agent workflow with the current human process, including how much ground each covers. Then compare the combined agent-and-review process with the same acceptance standard. Faster candidate generation can move the bottleneck into validation; if review capacity stays fixed, a larger pile of plausible outputs may lengthen the time to a trusted decision.

The OECD’s chapter on human and machine intelligence in science argues that knowledge systems still require domain curation and quality assurance, and that tools should be developed with the communities expected to use them. Although the chapter concerns science, the operating principle transfers: workflow fit and expert capacity belong in the investment case.

A decision gate for the next agent pilot

Before expanding a pilot, ask the owner to approve a short record containing the target decision, current baseline, coverage evidence, repeated-run results, validation method, cost per accepted result and the conditions that stop the trial. Record failures as evidence rather than deleting them from the demo deck.

For teams working across France and Morocco, one shared evaluation record can keep the engineering and business review aligned. It does not settle contractual, privacy or regulatory questions; those require evidence and approval for the specific engagement. Exceev’s services approach connects strategy, engineering and enablement, which are all needed when an agent result must become a process a team can operate.

Anthropic’s result is interesting because an agent noticed something that the original task did not ask it to find. The reruns are equally informative. Together they show a credible exploratory capability whose yield depends on orchestration, context and expert verification. That is enough reason to test similar workflows, and not enough reason to remove the people who decide what counts as proof.

Sources and limitations

We opened and reviewed every source below on 24 September 2026.

• Anthropic’s 23 September announcement establishes the announcement date, the company’s description of ART, the human laboratory role and the fact that ART’s function remains unknown.
• The Anthropic technical report provides the campaign design, task and token counts, laboratory findings, ten reruns and fixed-input benchmark results. It is a company preprint, not a peer-reviewed paper or independent replication.
• The NIST AI RMF Core supplies independent guidance on documented testing, verification, validation, oversight and review.
• The OECD chapter on collective and machine intelligence supplies independent context on expert curation, quality assurance and designing tools with their users.

We did not reproduce the campaign, inspect its private harness or data, price the model usage, review the biological evidence independently, or assess ART’s function. The report’s benchmarks use a model judge and configurations selected by Anthropic. No customer outcome or Exceev deployment is claimed. The proposed evaluation method is a business review, not scientific, legal or regulatory advice.