Anthropic’s research workflow filters hypotheses before lab testing

Anthropic describes a research workflow in which Claude searches public biological data, reproduces known results and writes evidence-backed candidate reports. Further analysis rejects most candidates before scientists choose experiments. The team reports identifying previously uncharacterized features around an enzyme system. The engineering lesson is the separation between generating many plausible hypotheses and gathering independent evidence that supports them.

Anthropic · Sep 23, 2026

Why it matters

A concrete agent workflow showing why a promising report is an intermediate result rather than a verified discovery.

More context

Read the workflow section for its staged review process. This is a vendor research account with a preprint, not a turnkey tool; the reported campaign used roughly 950 agents and human laboratory work.

Keep in mind

The system’s function remains unknown, and the underlying enzyme was previously identified. Human scientists performed the experiments. These early claims are not an independently reproduced evaluation.

Original read