Paperclip

The context layer under virtual-biotech: it digitises unstructured scientific material and maps scattered databases into one AI-native virtual file system, so agents reach millions of papers through ordinary file operations instead of per-database APIs. From Zou‘s Stanford group; documented here from stanford-virtual-biotech-37000-agents (T3), not from the repository.

The bet

It leans on what current models are already reliable at — writing code and walking a file system — rather than on their ability to drive a brittle query API. The stated problem is that MCP wrapping does not fix a bad interface: “the interface and APIs are not suitable for agents,” because they were designed for humans or pre-AI algorithms. A PDF dropped into a context window is the small version of the same failure, since text models misread figures and tables and then hallucinate.

Claimed effect: better accuracy, with “time and cost reduced by over an order of magnitude compared to if you use agents without these AI-native scientific infrastructures.” No benchmark, baseline or task set is given in the source.

Why it matters beyond biology

It is the same move as text-first-agent-design — give the agent a representation chosen for how agents work, not the one humans inherited — applied to data access rather than to a GUI or a document. If the finding generalises, the useful integration work at scale is rebuilding the interface, not adapting the agent to the interface that exists, which cuts against the prevailing enterprise approach of wrapping legacy stores in MCP servers.

Related

virtual-biotech · stanford-virtual-biotech-37000-agents · model-context-protocol · text-first-agent-design · agent-memory · james-zou · stanford-university