Limited published evidence · Scientific research
AI scientific literature review: FutureHouse example
FutureHouse runs literature search, evidence assessment, novelty review and report generation agents on Anthropic Claude within its Aviary platform. Crow, Falcon and Owl cover answers, deeper reports and novelty assessment. The source describes an early-stage platform, with Phoenix experimental.
Scientific literature search, evidence assessment, novelty review and report generation
FutureHouse
Published account: Implemented scientific-research platform and agents; Phoenix is experimental, and wider public access was still planned in the source.
A published third-party example, not a Nuvepro customer case or evidence of a partnership.
The workflow at a glance
FutureHouse reported work
Illustrative summary of the published account
Reported work components. Their arrangement does not imply a fixed execution order or simultaneous work.
- Research agents
Crow, Falcon and Owl
Cover literature answers, deeper reports and novelty assessment.
- Research tools
Specialist databases
Falcon accesses resources including OpenTargets.
- Reported evaluation
Ablations and LAB-Bench selection
The account describes model selection and effects of removing agentic steps.
What the source reports
Crow, Falcon and Owl cover literature answers, deeper reports and novelty assessment. Falcon accesses specialist databases including OpenTargets. The team reports controlled ablations in which removing agentic steps reduced answer accuracy, and describes LAB-Bench model selection. The case does not establish autonomous wet-lab operation.
Reported technology: Anthropic Claude within FutureHouse Aviary. Scope and outcomes are source-reported.
What the enterprise environment would need
The following is a proposed project design for an enterprise doing similar work. Published details are marked; the remaining systems, data and scenarios would be agreed for your project.
Proposed practice environment
Prepare a scientific evidence review
Illustrative project design to scope with your enterprise.
- 1 · Set up the environment
Research question corpus
Permitted literature snapshots, search results and contradictory studies form the evidence set.
- 2 · Preparation and agent contributions
Prepare evidence synthesis
Retrieve relevant studies and draft findings, citations and explicit uncertainty.
- 3 · Independent checks
Blind scientific review
Check citation support, recall and novelty claims against held-out expert judgments.
- 4 · People decide acceptance
Scientist reviews next steps
Accept the bounded research report and distinguish conclusions from possible experiments.
A reviewable evidence package
A sourced research report with conflicts and inaccessible evidence recorded.
If a check fails → withhold acceptance.
An inaccessible or contradictory study is silently replaced with unsupported confidence. Withhold acceptance, repair, independently recheck and return to owner review.
Failure or owner rejection returns to review
- Hold
- → 2 · Revise work
- → 3 · Rerun independent checks
- → 4 · Owner review
Operating boundary. Approved research tools only; no wet-lab execution or experiment authorization.
- Systems and tools
- Published: Aviary and specialist scientific databases. Proposed: reuse those tools; no full business application replica needed.
- Data and state
- Proposed: preserve literature corpus versions, search results, citations and permitted database access.
- People and handoffs
- Proposed: scientist judges evidence and selects subsequent experiments; evaluate that handoff separately from retrieval.
- Rules and authority
- Proposed: specify uncertainty and citation standards; chemistry restrictions need qualified safety review.
- Normal and exception cases
- Proposed: contradictory studies, inaccessible papers and no sufficient evidence.
- Reset and replay
- Proposed: replay fixed source snapshots while separately testing live retrieval drift; not a published operational feature here.
What needs checking
- Published: LAB-Bench selection and reported agentic ablation studies, without detailed results on this page.
- Proposed: blinded expert review, citation entailment and recall/precision on held-out questions.
- Proposed: use realistic research questions and access restrictions; generic office workflows would have weak fidelity.
Review environment fidelity, human capability and agent performance separately. The enterprise’s accountable owners decide acceptance of the shared workflow and any operating use.
Questions about this example
- What does FutureHouse use AI for?
- FutureHouse uses Claude within its Aviary platform for scientific literature search, evidence assessment, novelty review and report generation. Crow, Falcon and Owl cover literature answers, deeper reports and novelty assessment, and Falcon accesses specialist databases including OpenTargets.
- Who judges the evidence the FutureHouse agents produce?
- The source does not describe a review process for outputs. A proposed design, not FutureHouse's, has a scientist judge the evidence and choose follow-up experiments, evaluating that handoff separately from retrieval, with blinded expert review and citation checks.
- How strong is the evidence for FutureHouse's agents?
- The evidence is early-stage. The team reports controlled ablations where removing agentic steps reduced accuracy, and describes LAB-Bench model selection, but full ablation data is not supplied. Named enterprise adoption and production workflow acceptance are not shown, and Phoenix is experimental.
- What would an enterprise need to build something similar?
- This is Nuvepro's proposed design, not FutureHouse's. It would need versioned literature corpora, saved search results, citations and permitted database access. Tests would include contradictory studies, inaccessible papers and questions with no sufficient evidence, scored on held-out questions.
Shape this around your enterprise
Bring a similar workflow and the people responsible for it. Together, define the records and systems to use, the result your team will build or configure, and how its owners will check it. Nuvepro scopes the simulation, environment and preparation. Agree any existing-agent evaluation, implementation assistance and production responsibilities separately.
Sources and limitations
Curation review: . This is the date this source example was reviewed, separate from source publication dates.
- Named enterprise operational adoption, production workflow acceptance and full ablation data are not supplied.