{"resourceId":"anthropic-claude-agents-art-reverse-transcriptase-discovery-20261011","versions":[{"version":"external-0428d74b4c5bc87a05a1c468436ccf5bb195c0174e48feaf43cfabd61f5bde10","resource":{"id":"anthropic-claude-agents-art-reverse-transcriptase-discovery-20261011","title":"Anthropic preprint: Claude agents surveyed 1.9 billion protein clusters and flagged a new phage enzyme family, but ten reruns missed the key repeat array","organization":"Anthropic (Yoon, Athukoralage, Ameisen, Kauderer-Abrams, Perry, Durrant)","sector":"Industry research lab; relevant to university genomics and research computing","geography":"United States (San Francisco); computational genomics findings transfer broadly, lab work limited to BSL-1/BSL-2","publishedAt":"11 October 2026 (bioRxiv v1; technical report PDF hosted by Anthropic; trade coverage 17 October 2026)","publicationDate":"2026-10-11","eventDate":null,"sourceName":"Anthropic","sourceLabel":"Technical report / bioRxiv preprint (vendor-authored, not peer reviewed)","sourceUrl":"https://www-cdn.anthropic.com/22573675ada52a8ca8a97a1a4b4326b2f208a071.pdf","evidenceClass":"vendor-claim","outcomeClass":"mixed","topics":["developers-agents","knowledge-work"],"finding":"Anthropic ran an autonomous Claude Code harness (worker and supervisor agents with a shared record) on a written research brief. The campaign ran 119 tasks and 949 agent sessions, 77 agent-hours and 215.6 million tokens over 21.5 hours without human intervention, recovered about 200,000 RT clusters, and ended with 19 reports. Of 17 promoted candidate partner families, only three were confirmed as new RT associations; 14 were artifacts, known systems or bystanders. An agent spotted a repeat array by reading raw DNA, leading to a new family (ART). When the same campaign was run ten more times, the array was missed in every rerun. In fixed-input tests the strongest models described the array in at least 90% of attempts with DNA in the prompt but as low as 32% with files and tools. ART function remains unknown.","sledRelevance":"University genomics cores and research-computing centres are being pitched agent harnesses for data mining. This is a detailed, vendor-authored account of what such a campaign costs in sessions and tokens, how much human curation it still needs, and how unreliable a single run is.","evidence":"Opened the full technical report PDF on the Anthropic CDN (results, replicate-campaign methods, fixed-input benchmark) and the Anthropic research page; bioRxiv v1 date (11 October 2026) confirmed through the bioRxiv details API. The Drug Discovery News article of 17 October returned HTTP 403 and was not used.","architectureImplications":"Multi-agent harnesses need a shared task record, supervisor review and full session transcripts; model input design (DNA in context versus files) changed recognition rates sharply.","governanceImplications":"Single-run discoveries are not reproducible by default; the authors needed replicate campaigns, fixed-input benchmarks and human review to judge the result. Agent outputs should be treated as candidate reports for expert curation.","securityPrivacyImplications":"Not addressed for data access; the lab reports work only at BSL-1/BSL-2 with no human pathogens, and all wet-lab work by human scientists.","caveats":"Vendor-authored preprint about its own models, not peer reviewed; model versions used are proprietary; dollar cost not reported; biological function of ART not established; replicate search could miss ART loci on contigs outside the original identifier set.","streamIds":["research"],"roles":{"sales":"Lead with the reproducibility gap. Anthropic's own report shows an agent campaign that found a new enzyme family once, then missed the key feature in ten reruns, with 14 of 17 candidates rejected. Genomics cores and research-computing leaders evaluating agent pilots should hear that agents generate candidates for expert review, not finished discoveries. Interpretation: this is a vendor account, so it supports a bounded pilot design, not claims of faster discovery or cost savings.","engineering":"Design for repeat runs and inspectable inputs. The harness used worker and supervisor agents, a shared record and full transcripts; recognition of the repeat array depended on whether raw DNA entered the model context, dropping to as low as 32% with files and tools. A pre-sales design should include replicate campaigns, transcript retention and token metering. Interpretation; model versions are proprietary and costs in dollars are not reported.","delivery":"Frame delivery as a curation pipeline around agents. The campaign produced 19 reports, and humans plus replicate runs and benchmarks decided what held up. Delivery would set briefs, review criteria, replicate-run schedules and a path to lab validation, owned by the core facility lead. Interpretation: acceptance should be measured by curated candidates that survive review and replication, not by agent output volume."},"roleDetails":{"sales":{"customerProblem":"Research groups want to mine growing sequence databases faster but cannot rely on a single agent run or staff enough curators.","stakeholders":["Genomics or bioinformatics core director","Research computing director","Principal investigators in microbiology","Research office"],"discoveryQuestions":["How would you check that an agent finding reproduces across runs?","Who curates agent-generated candidate reports, and at what volume?","Do you meter and budget tokens and agent sessions per project?"],"valueHypothesis":"A bounded agent pilot with replicate runs may surface candidates for expert review that manual surveys miss, at a measurable token cost.","potentialEngagement":"A pilot design study for one agent-assisted data-mining campaign, including replicate-run protocol, transcript retention and curation criteria.","unsupportedClaims":["Agents reliably discover new biology in a single run","Guaranteed time or cost savings versus expert curation","Dollar cost estimates (the report gives tokens, not dollars)"]},"engineering":{"fit":"Fits groups with large sequence or data archives and existing HPC; limited by proprietary model access and reproducibility.","architectureIntegration":"Worker and supervisor agents with a shared task record on campus compute or cloud, with full transcript and token logging.","prerequisites":["Access to a capable agent model and harness","Curated reference data and profile models","Storage for transcripts and task records"],"constraints":["Results varied across reruns of the same brief","Proprietary model versions limit independent replication","Sensitive or controlled data may not be allowed in external model services"],"security":"The source does not cover data access controls; keep agent tool permissions and data egress under institutional review.","proofOfValue":["Run the same brief at least several times and compare findings","Benchmark recognition with inputs in context versus files","Measure the share of agent candidates that survive expert curation"]},"delivery":{"work":"Set up briefs, replicate-run protocol, curation rubric and handoff to lab validation for agent-assisted mining.","dependencies":["Agent platform contract or access","Domain experts for curation","Wet-lab capacity for validation"],"ownership":"Core facility or bioinformatics lead, with PIs accountable for scientific claims.","skillsAdoption":"Curators learn to read agent transcripts and reports; staff learn to write research briefs and manage token budgets.","governanceCheckpoints":["Replicate-run review before any candidate is reported","Expert sign-off on each report","Disclosure of agent use in publications"],"acceptanceCriteria":["Proposal: every reported candidate reproduced in at least one replicate run or fixed-input test","Proposal: token and session use logged per campaign","Proposal: curation decisions recorded for all agent reports"],"risks":["Overstating single-run findings","Curator overload from report volume","Dependence on one vendor model"]}},"retrievedAt":"2026-10-18T21:59:00-05:00","enrichedAt":"2026-10-18T21:59:00-05:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Shifts expert effort from scanning data to reviewing many agent reports and designing briefs; Anthropic says hypothesis volume itself became an object of study.","procurementImplications":"Buyers should ask for repeat-run reproducibility, token and session consumption, and the share of agent candidates that survive expert review, not single headline discoveries.","operatingModelImplications":"Agent campaigns work as a candidate-generation layer feeding human curation and lab validation, with replicate runs as a standard step.","sourceVerification":{"openedUrl":"https://www-cdn.anthropic.com/22573675ada52a8ca8a97a1a4b4326b2f208a071.pdf","referenceExcerpt":"the array was missed in every rerun","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}