Autonomous AI scientists

Named AI systems that take meaningful initiative in hypothesis generation, experiment design, or analysis. Two recent peer-reviewed Nature papers (Google’s Co-Scientist, FutureHouse’s Robin) anchor the wet-lab end of the landscape; general-purpose biomedical agents (Biomni, CRISPR-GPT) and chemistry agents (Coscientist, ChemCrow) cover the rest of the field.

This section is informational — a map of the field for reference and comparison. These are external systems, not components the Composer installs or runs; instead, the Composer can assemble an equivalent multi-agent system from cataloged Skills and MCP servers. Where a system also ships as an installable Claude component, it additionally appears in the Catalog (e.g. Biomni as a Claude Skill).

Where to start

  • Landscape — what’s currently working, how systems are evaluated, and the open problems.
  • Systems — one page per named system, with affiliation, autonomy level, validation, availability, and citations.

What “autonomous” means here

Autonomy here is a spectrum:

  • Assistive — human-in-the-loop, agent suggests, human decides.
  • Semi-autonomous — closed-loop with checkpoints.
  • Fully autonomous — closed-loop without intervention.

The three primary lifecycle stages are hypothesis generation, experiment design, and analysis. A system may cover one stage, several, or close the full loop.

See also

  • Contribute — file a recipe question if you want to use one of these systems on a concrete problem and aren’t sure how.

Table of contents