Presents ANTMAN, an adaptive coordination framework that organizes multi-agent information-seeking around evolving unresolved query needs rather than static partitions.
ANTMAN is an adaptive coordination framework that uses a revisable Need Graph to track unresolved requirements, accumulated evidence, and search progress, allowing it to dynamically control worker selection, routing, and task-local recovery. By decoupling coordination from the size of the information space, ANTMAN increases active coordination by only 1.23× under a 16× increase in searchable context, compared to more than 15× for partition-driven baselines, while maintaining strong answer quality. This approach ensures that computation scales with query demand rather than the segmentation of the underlying data, enabling effective navigation across multi-document question answering, long-context scaling, and structured environments.
ANTMAN addresses a central coordination problem in multi-agent information-seeking: how should several agents navigate a large, heterogeneous information space when the relevant subproblems are not known in advance? Rather than assigning agents to fixed slices of a corpus, topic, or query decomposition, the work frames coordination around evolving unresolved query needs. The core premise is that in complex search, research, or knowledge-navigation tasks, the useful division of labor is not static; it should change as agents encounter evidence, contradictions, dead ends, and newly relevant sources. ANTMAN therefore treats the unresolved needs of the query—what is still missing, uncertain, conflicting, or insufficiently grounded—as the primary objects that agents track, prioritize, and act upon.
The key contribution is a coordination paradigm in which multi-agent navigation is organized by a shared, dynamic model of need states rather than by rigid role or partition assignments. Agents do not merely search independently and later merge results; they operate with respect to the current set of open needs, selecting navigation actions that are expected to reduce uncertainty or fill specific gaps. This makes the system more adaptable to long-horizon, cross-source queries where early discoveries can invalidate or reshape the original plan. The insight is that effective multi-agent information navigation depends less on how the workspace is initially divided and more on how well the system can maintain a live account of what remains unresolved and which agent or action is best positioned to address it next.
This matters because many modern multi-agent search and retrieval systems still rely on static decomposition, fixed tool pipelines, or loosely coupled agent ensembles that can become brittle in large information spaces. ANTMAN is relevant to LLM-based research assistants, enterprise knowledge discovery, scientific literature navigation, and other settings where queries are multi-hop, cross-domain, and sensitive to newly discovered context. Its significance is architectural as much as algorithmic: it shifts the design focus from partitioning the information space in advance to coordinating agents around the evolving epistemic state of the task, which can reduce redundant effort, improve coverage, and support more robust adaptation as the task unfolds.