watch agent run

An observability layer for agent runs: a repository that turns opaque agent logs into deterministic readouts — input/output tables, color-coded views and a provenance report — so scientists can compare what an agent claims it did with what append-only logs show actually happened. Demonstrated on macrophage differentiation protocols extracted from papers, where three prompting approaches looked identical at eight papers but diverged sharply at 200: every run claimed full coverage, and only one achieved it. What you can see, you can begin to trust.

See the code →