Claude Science is Anthropic's beta AI workbench for scientific research. It packages Claude as a coordinating agent inside a research environment that connects to scientific databases, packages, local or remote compute, and domain-specific skills. Unlike a general chat assistant, the product emphasizes auditable artifacts: figures, manuscripts, code, environment details, and message history are kept together so results can be validated and reproduced later.^[raw/articles/claude-science-ai-workbench-2026.md]
The product is especially relevant to ai-research-automation because it turns research work into a persistent, tool-connected loop rather than a one-off answer. Claude Science can run locally on macOS/Linux or against remote machines over SSH/HPC login nodes, ask before reaching new resources, submit jobs, and fork sessions to compare approaches. A reviewer agent checks citations, calculations, untraceable numbers, and figure/code consistency, which connects the product to ai-evaluation-infrastructure and loop-engineering.
Design signals
Auditable artifacts: outputs include the code and environment that produced them, plus plain-language explanations and message history.
Compute as part of the loop: local machines, lab infrastructure, HPC, and Modal-style on-demand compute are treated as execution targets behind explicit review/revoke decisions.
Domain skills and connectors: the beta ships with 60+ curated skills/connectors for areas such as genomics, single-cell, proteomics, structural biology, and cheminformatics.
Reviewer agents: separate critic/reviewer agents are used to catch citation and calculation errors, echoing the generator/evaluator split in loop-engineering.
Open Questions
How much of the reproducibility guarantee depends on preserving exact execution environments versus preserving narrative provenance?
Can the reviewer-agent pattern transfer to personal wiki curation, code review, and scheduled research jobs without becoming too expensive?
What privacy boundary is acceptable when sensitive lab data remains local but context is still sent to a hosted model?
