SearchSwarm: Towards Delegation Intelligence in Agentic LLMs for Long-Horizon Deep Research

Researchers present SearchSwarm, a framework that trains large language models to intelligently delegate complex tasks to subagents while managing finite context windows. The resulting 30B-parameter model achieves state-of-the-art performance on research benchmarks by learning when and what to delegate, addressing a critical bottleneck in agentic AI systems.

SearchSwarm tackles a fundamental constraint in deploying large language models for complex, multi-step reasoning tasks: context window limitations. As models handle increasingly sophisticated real-world problems, the amount of information needed grows exponentially, but the fixed context window remains a hard constraint. The paper's innovation lies not in expanding context windows but in teaching models strategic delegation—deciding which subtasks to outsource to specialized agents and how to integrate their results.

The delegation problem has grown more acute as AI systems move toward agentic architectures where a main agent orchestrates multiple specialized workers. Previous approaches lacked principled methods for teaching models this capability since delegation patterns rarely appear in naturally occurring text. SearchSwarm solves this by creating a harness that guides task decomposition and constrains subagent outputs, generating high-quality supervised training data. This synthetic data approach mirrors recent trends in AI where engineered training signals replace scarce natural examples.

The performance gains are substantial. SearchSwarm-30B-A3B outperforms comparably-sized models on BrowseComp benchmarks, suggesting that delegation intelligence is a learnable and valuable capability. For developers building research and analysis tools, this demonstrates that model scale isn't the only lever for improving long-horizon task performance. The promise to release the harness, weights, and training data could accelerate industry adoption of delegative agent patterns.

The broader implications extend to resource efficiency in AI systems. Rather than continuously scaling context windows or model parameters—both computationally expensive—teaching models to delegate enables systems to handle complex problems with constrained resources. This efficiency gain matters for both cost-conscious deployments and environmentally conscious AI development.

• →SearchSwarm introduces delegation intelligence through supervised fine-tuning on harness-guided trajectories, enabling models to strategically decompose and delegate complex tasks.
• →The 30B-parameter model achieves 68.1 on BrowseComp and 73.3 on BrowseComp-ZH, outperforming comparable models by learning when to delegate subtasks.
• →Synthetic training data generated from structured harnesses can effectively teach agentic behaviors that rarely appear in naturally occurring text.
• →Delegation-based architectures offer a more efficient alternative to context window expansion for handling long-horizon reasoning tasks.
• →Open-source release of harness, weights, and training data positions SearchSwarm as foundational infrastructure for agentic AI development.