SearchSwarm: Empowering Agents with Delegation Capabilities for Long-Range Deep Research

This article introduces SearchSwarm—an agent solution for long-range deep research tasks. Through a harness-guided training method, it internalizes delegation intelligence into model capabilities, enabling a 30B-parameter model to achieve a score of 68.1 on the BrowseComp benchmark. It also open-sources the harness, model weights, and training data to facilitate the development of multi-agent systems.

Large language models have limited context windows, while long-range deep research tasks require processing large amounts of search, information integration, and cross-document reasoning, which exceed the capacity of a single model. To resolve this contradiction, delegation intelligence is needed: the main agent decomposes subtasks to sub-agents, and sub-agents return summaries to save context. Delegation intelligence includes three core capabilities—task decomposition, delegation decision-making, and result integration—but the lack of explicit delegation records in natural text leads to a scarcity of training data.

SearchSwarm designs a harness constraint framework to guide the main agent to decompose subtasks reasonably and sub-agents to return results in the required format. It uses the trajectories generated by the harness as supervised fine-tuning data to internalize delegation intelligence into model weights. This method does not require expensive human annotation and can generate high-quality training data at scale.

The SearchSwarm-30B-A3B model achieved a score of 68.1 on the BrowseComp benchmark and 73.3 on BrowseComp-ZH, both being the best among models of the same scale. These results were achieved by a model with only 30 billion active parameters (A3B), demonstrating the potential of an efficient architecture.

The research team open-sourced the harness, model weights, and training data, providing a foundation for the community to research delegation intelligence. For developers, SearchSwarm offers a reproducible path: generating training data through a constraint framework, internalizing complex capabilities into the model itself, simplifying the architecture, and improving end-to-end efficiency and reliability.