Virtual Biotech teardown: 6 design decisions behind Science's 37,000-agent drug company

bravo_abad · x · 2026-09-22

A detailed systems teardown of Zhang et al.'s Virtual Biotech (Science, Sept 2026): 10 MCP servers with 3–16 tools each, a two-algorithm/four-phase orchestration flow, typed tool contracts, schema-validated outputs, self-run statistical checks, and a read-only reviewer on the cheapest model that still catches errors. The author argues these implementation choices — not the org chart of a virtual CSO and eight AI scientists — are what make the system's science auditable, and that most principles transfer beyond biomedicine.

Related event: Stanford's VirtualBiotech: Tens of Thousands of AI Agents Run a Virtual Pharma, Published in Science(3 posts)→

Original post →

More from coding & agent

coding & agent channel →