A configurable and hierarchical tree-based static Deep Research Agent designed to overcome the limitations of standard RAG pipelines for complex, multi-turn research tasks.

The Static Deep Research Agent (Static-DRA) introduces a novel approach to autonomous research by utilizing a static workflow governed by two user-tunable parameters: Depth and Breadth. Unlike dynamic workflows that can spiral in cost and complexity, Static-DRA allows users to explicitly balance the desired quality/comprehensiveness of the research against the computational cost (LLM tokens).

This architecture employs a hierarchical team of agents—Supervisor, Independent, and Worker—to facilitate multi-hop information retrieval and parallel sub-topic investigation.

The system operates on a hierarchical tree structure where a primary research topic is decomposed into sub-topics based on the configured parameters.

1. Supervisor Agent: Determines if a research topic can/should be split further. It consults the LLM to decide if the current node requires decomposition or direct research.

2. Independent Agent: Acts as a middle-manager. If a topic is split, it spawns child Supervisor agents for every sub-topic and aggregates their results.

3. Worker Agent: The leaf node of the tree. It performs the actual deep research using:

   • LLM (Reasoning): Currently evaluated with gemini-2.5-pro.

   • Web Search Tool: Uses Tavily API to fetch relevant citations (filtering for relevance scores > 30%).

The expansion of the research tree is controlled by the Depth (d) and Breadth (b) parameters.

• Depth: How many levels deep the agent recursively breaks down topics.

• Breadth: The maximum number of sub-topics spawned from a single parent topic.

Below are the evaluation metrics for Static-DRA at multiple configurations.

Note: The effective breadth decreases by a factor of 2 at each deeper level to focus resources on the most relevant core topics.

We evaluated Static-DRA against the DeepResearch Bench using the RACE (Reference-based Adaptive Criteria-driven Evaluation) framework.

Configuration:

• Model: gemini-2.5-pro
• Depth: 2
• Breadth: 5

If you refer or use Static Deep Research Agent (Static-DRA) in your research, please cite my paper:

@misc{prateek2025hierarchicaltreebasedapproachcreating,
      title={A Hierarchical Tree-based approach for creating Configurable and Static Deep Research Agent (Static-DRA)}, 
      author={Saurav Prateek},
      year={2025},
      eprint={2512.03887},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2512.03887}, 
}