{"resourceId":"fujimura-autonomous-research-agents-formulate-own-questions-20261029","versions":[{"version":"external-206f8af3f42c0b62170fd161f8b94df301d295c4515b9926f670b08611a15046","resource":{"id":"fujimura-autonomous-research-agents-formulate-own-questions-20261029","title":"Preprint: a coding agent chose its own research questions in five pre-registered runs, all producing re-computable papers rated near workshop level, with low novelty","organization":"Ko Fujimura, Otsuma Women's University (single-author arXiv preprint)","sector":"Academic research on autonomous AI research agents","geography":"Japan","publishedAt":"Submitted 29 October 2026 (arXiv v1)","publicationDate":"2026-10-29","eventDate":null,"sourceName":"arXiv","sourceLabel":"Fujimura, \"Letting AI Formulate the Research Question: Autonomous Research Agents,\" arXiv:2610.39620, 29 Oct 2026","sourceUrl":"https://arxiv.org/abs/2610.39620","evidenceClass":"academic-research","outcomeClass":"mixed","topics":["developers-agents","governance-procurement","knowledge-work"],"finding":"An off-the-shelf coding agent, managed by Claude Code with Claude Opus 5.5, was given a roughly 7,300-word instruction document, a question-formulation skill (with one critique from gpt-6-astra under a US$20 budget) and a peer-review skill in which a reviewer agent reruns the study from its records. After 13 development runs, the system was frozen and five consecutive runs (one pilot counted by disclosed amendment, four pre-registered) were all reported. All five went from question to a revised, re-computable paper on a laptop with no human research input; external AI reviewers rated them at or near workshop level, but novelty and generality remained low. Small wording changes to the research-area setting shifted topics substantially. Records include claims-to-evidence logs, API costs, Git tags and a verification entry point.","sledRelevance":"Relevant to research offices setting policy on agent-generated papers: the study shows end-to-end agent research is feasible on modest hardware and that full logs and pre-registration make it auditable, while also showing low novelty and AI-only evaluation. It informs disclosure, authorship and review rules.","evidence":"Read the abstract, introduction, design, discussion and limitations. Five runs with one research model, no ablation, AI-only evaluation (human evaluation planned), in-loop reviewer from the same model family, and post-hoc choices disclosed as deviations. The paper itself was drafted by the management agent. Preprint, not peer reviewed.","architectureImplications":"Management agent, research agent and reviewer agent in separate workspaces; local laptop computation; one budgeted API call to a different model for critique.","governanceImplications":"Pre-registration, reporting every run and a claims-evidence record are concrete controls institutions could require for agent-produced work.","securityPrivacyImplications":"Isolation between runs relied on configuration and instructions, not access control, a weak control for sensitive settings.","caveats":"Very small sample, single author, AI-only ratings uncalibrated, no human evaluation yet; one research area setting tested. Non-US source; transfer to US institutional settings depends on local policy.","streamIds":["research"],"roles":{"sales":"Lead with governance for agent-produced research. A 29 October preprint reported five pre-registered runs in which a coding agent chose its own question and produced a re-computable paper on a laptop; AI reviewers rated them near workshop level with low novelty. Research integrity officers and graduate deans need disclosure and authorship rules. Offer a bounded review of policy on agent-generated manuscripts and required records. Do not claim agents produce publishable research; evaluation was AI-only and small.","engineering":"Frame fit around auditable agent pipelines. The system separates management, research and reviewer agents, logs each claim with its evidence, records API calls and costs, tags drafts in Git and offers a verification entry point that regenerates numbers without AI calls. Prerequisites are sandboxed workspaces, version control and logging. Validate by independently re-implementing a sample analysis rather than rerunning the agent's own code, and by enforcing run isolation with access controls rather than instructions.","delivery":"Treat delivery as a policy and tooling rollout. Work covers disclosure rules for agent-produced work, a records standard (pre-registration, all runs reported, claims-evidence logs) and reviewer training. Ownership sits with the research integrity office with graduate school and library input. Governance checkpoints cover human review before submission and authorship statements. Acceptance criteria are proposals, such as every agent-assisted manuscript carrying a verifiable records package, not results the author reported."},"roleDetails":{"sales":{"customerProblem":"Researchers can now generate whole papers with agents, and institutions lack rules for disclosure, records and authorship.","stakeholders":["Research integrity officer","Graduate school dean","Vice president for research","Library scholarly communications"],"discoveryQuestions":["Do current AI policies cover fully agent-run studies?","What records must accompany AI-assisted manuscripts?","Who reviews agent-produced work before submission?"],"valueHypothesis":"A records standard may let institutions permit agent research while keeping it auditable.","potentialEngagement":"A review of agent-generated manuscript policy and required records package.","unsupportedClaims":["Agents produce publishable research","AI reviewer scores reflect quality","Results generalize beyond five runs"]},"engineering":{"fit":"Fits computational studies run by coding agents with full logging.","architectureIntegration":"Separate management, research and reviewer agents; Git-tagged drafts; claims-evidence logs; AI-free verification entry point.","prerequisites":["Sandboxed workspaces","Version control","Run logging and cost records"],"constraints":["Isolation enforced only by instructions","Same-family reviewer","Computational topics only"],"security":"Enforce run isolation and data access with permissions, not prompts.","proofOfValue":["Independent reimplementation of a sample analysis","Human expert rating of outputs","Verification script regenerates all numbers"]},"delivery":{"work":"Write disclosure policy, define records package, train reviewers and pilot with volunteer labs.","dependencies":["Existing AI use policy","Repository and logging tools","Faculty reviewers"],"ownership":"Research integrity office with graduate school and library.","skillsAdoption":"Researchers learn pre-registration and records practices for agent work.","governanceCheckpoints":["Human review before submission","Authorship and AI-involvement statement","Pre-registration check"],"acceptanceCriteria":["Each agent-assisted manuscript has a verifiable records package (proposal)","Disclosure rate tracked","Reviewer training completed"],"risks":["Low-novelty paper flood","Unverified AI evaluations","Isolation failures between runs"]}},"retrievedAt":"2026-11-02T22:07:31-06:00","enrichedAt":"2026-11-02T22:07:32-06:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Suggests the scarce human skill is choosing questions that matter, which the agent did poorly.","procurementImplications":"Uses commercial model subscriptions and small API budgets; no special infrastructure.","operatingModelImplications":"Agent-run studies need records and review built in; reviewers rerunning code is not independent implementation.","peerReview":"preprint","sourceVerification":{"openedUrl":"https://arxiv.org/abs/2610.39620","referenceExcerpt":"All five runs went from question formulation to a revised, re-computable paper, with local computation on a laptop and no research input from humans","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}