Shows where agents break on long dependent workflows (dropped qualifiers, evidence-discarding shortcuts) and that expert plans and extra thinking time raise scores, which guides how to scaffold agents for multi-step domain work.
Key takeaways · AI-distilled
DrugDiscoveryBench, built by Scale Labs and Phylo, has 82 expert-written problems covering molecular structure lookup, chemical property calculation, and patent and genetic database mining. The strongest agents solved about half.
The leaders, all from OpenAI, Anthropic and Google, clustered within a few points of one another; agents from other companies trailed far behind, and a handful of problems stumped every AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → tested.
Agents broke on tasks with seven or eight dependent steps: a qualifier dropped mid-task, a shortcut discarded required evidence, or the agent worked on the wrong subject, then submitted a confident wrong answer.
Given an expert's step-by-step plan that did not reveal the answer, tasks that scored zero unaided reached perfect or near-perfect scores, while the agent still ran every query and calculation itself.
The surrounding system mattered: some agents gained as much as 17 points just by switching agent harnessThe scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.Full definition →, and the same agent given more time to think improved by close to 20 points.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
