An AI agent built to run drug-discovery projects can't out-think the chemistry models it depends on.
Researchers built MAGI, an open-source agent that sets its own objectives, runs molecule-optimization loops, reads structure-activity data, and revises its strategy as it goes. It generates candidate molecules either directly through an LLM or by delegating to REINVENT 4, with the scoring tools swappable behind a shared interface. The team tested it on nine real lead-optimization campaigns pulled from three pharmaceutical companies, replaying each one with only the data available at the time. Both generation routes produced valid molecules: the LLM route stuck closer to known chemistry and hit target activity in fewer steps, while REINVENT ranged further into unfamiliar chemical space. But whether a campaign actually succeeded tracked the accuracy of the underlying prediction models on the chemistry proposed, not which generation method was used.
In a blinded test, chemists could not tell MAGI's proposals from real, held-out compounds, and judged its structure-activity reasoning plausible, if incomplete. That's a strong showing for an open, pluggable system, but it also relocates the real bottleneck in agentic drug discovery: not how well the agent plans or explains itself, but how good the scoring models are once the chemistry drifts outside what they've seen.
Call it the field's version of running out of training data - the agent coordinates the work just fine, it's the yardstick measuring it that gives out first.