Can a shared world model sustain coherence across hundreds of agent steps?

Kosmos is presented as an AI scientist that stays coherent over far more steps than its predecessors, and the paper credits one mechanism: "a structured world model to share information between a data analysis agent and a literature search agent." The abstract reports runs of "up to 12 hours" across "200 agent rollouts," averaging "42,000 lines of code" and "1,500 papers" per run. The most checkable figure is reliability: "Independent scientists found 79.4% of statements in Kosmos reports to be accurate." These are the builders' report of their own system, so they are claims to weigh, not outside benchmarks.

The mechanism is a loop with shared state. Each cycle, Kosmos runs "up to ten literature search and analysis tasks," writes summaries of their outputs into the world model, and queries that model to propose the next cycle's tasks. The authors frame this as context management that lets the system "explore many different research avenues simultaneously." The same structure supports traceability, since each statement in a report cites "a publication found by the literature search agent or a Jupyter notebook created by the data analysis agent." The Discussion extends the design to any field and shows results in six: metabolomics, materials science, connectomics, statistical genetics, proteomics and transcriptomics.

Coordination here is centralized. One world model holds the shared state and decides what runs next, which cuts against the design in Can decentralized teams outperform central planners in long-running science?. The contrast is partial, since Kosmos also works from an objective a scientist sets. Can decentralized agents coordinate research without a central planner? shares results with no planner at all, while Kosmos keeps a summarized, continuously updated state. The nearest mechanism is in Can AI research itself without losing human oversight?, where both systems turn outcomes into a store the next step reads from. Kosmos's store is filled by the run's own outputs, whereas Can AI automate the discovery of how AI models work? draws on an existing knowledge graph.

The excerpt does not test whether the world model causes the coherence; it reports no run without one. The accuracy figures are also uneven. The 79.4% is attributed to independent scientists, but the Limitations section reports the authors' own evaluations: 85% of statements from data analyses were accurate, against 57% of statements that "required interpretation of results." The "6 months" equivalence and the linear scaling of findings with cycles come from collaborators, not from counts in the paper, and the authors note that identifying valuable discoveries "is a time-intensive process that relies on human scientists." At the strength the evidence allows, Kosmos is supported as a high-volume, traceable generator of analyses and hypotheses whose interpretive claims need expert review. The authors say as much: "the central value proposition is therefore not that Kosmos is always correct."