Static Deep Research Agent (Static-DRA)

• Static-DRA is a fixed, tree-structured autonomous research agent designed to execute complex multi-turn tasks using a pre-planned research agenda.
• It features configurable depth and breadth parameters, enabling precise trade-offs between comprehensive research output and computational cost.
• Its three-agent architecture—Supervisor, Independent, and Worker—ensures modular execution, predictable debugging, and strict workflow compliance.

A Static Deep Research Agent (Static-DRA) is an autonomous research agent operating over a fully pre-specified, static, tree-structured workflow for complex, multi-turn research tasks. In contrast to dynamic agents that interleave real-time LLM-based planning and reflexive adaptation, Static-DRA executes a fixed agenda from information acquisition through to analytical synthesis and structured report generation. This design provides rigorous user control over both research intensity and computational cost, while imposing a strict separation between workflow structure and run-time content (Prateek, 3 Dec 2025, Huang et al., 22 Jun 2025).

1. Hierarchical Tree-based Workflow and Static Planning

Static-DRA organizes the research process as a static, hierarchical tree. The root node corresponds to the original research topic. Each internal node (handled by a Supervisor and Independent agent) attempts to decompose its assigned topic into up to $b$ sub-topics, while leaf nodes (handled by Worker agents) are responsible for final LLM-aided and web-based research. The recursive split procedure terminates either upon reaching the configured depth $d$ or when the LLM determines that further subdivision is not possible. Critically, the entire tree structure is instantiated in a single top-down pass before any downstream modules are invoked, enforcing a static workflow:

$\begin{tikzpicture}[level distance=1.2cm, sibling distance=1.8cm] \node (R) [draw, rectangle] {Root: Research Topic} child { node [draw, rectangle] {Subtopic 1} child { node [draw, ellipse] {Worker: Research 1.1} } child { node [draw, ellipse] {Worker: Research 1.2} } child { node [draw, ellipse] {Worker: Research 1.3} } } child { node [draw, rectangle] {Subtopic 2} child { node [draw, ellipse] {Worker: Research 2.1} } child { node [draw, ellipse] {Worker: Research 2.2} } child { node [draw, ellipse] {Worker: Research 2.3} } } child { node [draw, rectangle] {Subtopic 3} child { node [draw, ellipse] {Worker: Research 3.1} } child { node [draw, ellipse] {Worker: Research 3.2} } child { node [draw, ellipse] {Worker: Research 3.3} } }; \end{tikzpicture}$

Such a static research tree guarantees deterministic execution, no mid-run re-planning, and facilitates module isolation for debugging and compliance (Prateek, 3 Dec 2025, Huang et al., 22 Jun 2025).

2. Depth and Breadth Configurability

Static-DRA introduces two user-configurable hyperparameters:

• Depth ($d$): Maximum tree depth; when a node reaches $d=0$, it is routed directly to Worker.
• Breadth ($b$): Maximum number of children an internal node may generate.

Assuming no pruning, the research graph contains: $N = \sum_{i=0}^{d} b^i = \frac{b^{d+1} - 1}{b-1}$ nodes (Supervisor, Independent, Worker roles). Each node, typically invoking a single LLM call at cost $c_\mathrm{LLM}$, yields a total resource cost $C = N \times c_\mathrm{LLM}$.

In practice, Static-DRA applies dynamic breadth reduction by decrementing $b$ by 2 per level; the count of leaf topics is

$ns(d) = \prod_{i=0}^{d-1} \max(b-2i, 1)$

Larger $d$ or $b$ yield greater report comprehensiveness at increased computational cost, enabling explicit research/cost trade-offs (Prateek, 3 Dec 2025).

3. Agent Architecture: Supervisor, Independent, Worker

Static-DRA operationalizes a three-agent hierarchy:

• Supervisor Agent: Given $(\text{topic}, d)$, queries the LLM whether the topic can be split. If so, invokes Independent agent on $(\text{topic}, d-1, b-2)$; else (if topic is unique) invokes Worker on $(\text{topic})$.
• Independent Agent: Receives $(\text{topic}, d, b)$, triggers the LLM to generate up to $b$ sub-queries, spawns Supervisors for each sub-query in parallel, and merges sub-reports.
• Worker Agent: Receives a leaf topic, uses web search (filtered ≥30% relevance), forwards selected URLs and topic to LLM for final summary and citations.

This architecture implements multi-hop retrieval and parallel sub-topic investigation, facilitating breakdown of complex queries across research sub-domains (Prateek, 3 Dec 2025). Pseudocode for these agents is specified in the source (Prateek, 3 Dec 2025).

4. Workflow Example and Data Flow

In a full execution, user input flows through static planning, fixed tool invocation, and sequential phase updates:

Phase	Operation	Agent/Module
Information Acquire	Domain-specific APIs / browser scrape	Worker
Static Planning	Subtopic decomposition / agenda fill	Supervisor, Independent
Analysis/Synthesis	Predefined analytics, code execution	Worker (scripted)
Report Generation	Structured LLM template fill	Worker

No phase is adaptively altered based on intermediate outcomes; all branching and analysis are determined up-front and executed without LLM-driven re-optimization (Prateek, 3 Dec 2025, Huang et al., 22 Jun 2025).

5. Benchmarking and Empirical Analysis

Static-DRA is evaluated via DeepResearch Bench, encompassing 100 PhD-level tasks across 22 fields, using the RACE (Reference-based Adaptive Criteria-driven Evaluation) framework. RACE scores on five axes—Comprehensiveness, Insight, Instruction-Following, Readability, Overall (mean):

• Example configuration $d=2$, $b=5$, model: gemini-2.5-pro:
  • Overall = 34.72
  • Comprehensiveness = 35.12
  • Insight = 30.45
  • Instruction-following = 38.86
  • Readability = 35.44

Smaller $(d, b)$ yield lower scores and lower LLM usage; e.g., $d=1$, $b=2$ yields $\sim$22; $d=2$, $b=3$ yields $\sim$28 (Prateek, 3 Dec 2025). The score rises monotonically with LLM call count, quantifying the cost-comprehensiveness trade-off.

A secondary FACT framework assesses factual retrieval/citation quality, though details are not treated in depth in (Prateek, 3 Dec 2025).

6. Relationship to General Static Deep Research Architectures

Static-DRA is a specialized instance of the broader Static-Deep Research Agent paradigm (Huang et al., 22 Jun 2025). In this paradigm, the agent’s workflow—data acquisition, planning, tool use, and report synthesis—is hardcoded, with fixed phase boundaries and no run-time LLM feedback. Static-DRA distinguishes itself with a configurable tree topology, whereas classical static pipelines (e.g., AI Scientist, Agent Laboratory) define linear or fixed-phase execution. Static workflows emphasize decomposability, modularity, predictability, and phase-wise isolation, and are contrasted with dynamic workflows that employ run-time LLM-based task adaptation and flexible subtask orchestration (Huang et al., 22 Jun 2025).

7. Practical Implications, Limitations, and Best Practices

Static-DRA’s explicit $(d, b)$ configurability empowers users to balance research thoroughness and compute costs, aligning process granularity with task scope and resource constraints. The open-source codebase exposes modular agent implementations, web-search clients, structured prompts, benchmark scripts, and output reports (Prateek, 3 Dec 2025). The LLM backend is swap-compatible.

Advantages: Predictable, efficient, debuggable, auditable, well-suited to well-scoped tasks with fixed informational needs (Huang et al., 22 Jun 2025).

Limitations: Incapable of run-time adaptation—if sources, analytic requirements, or evaluation criteria change, the workflow cannot react without manual modification; performance degrades on unstructured or open-ended research queries.

For robust deployment, best practices include phase modularization with clear input/output schemas, balanced memory summaries to control context length, extensive offline testing, and phase-wise as well as aggregate metric evaluation. For domains with partial structure, hybrid “semi-static” workflows may be appropriate, merging fixed pipelines with occasional dynamic recalibration (Huang et al., 22 Jun 2025).

A plausible implication is that as research workflows and evaluation tasks become more open-ended, dynamic agents will supersede static architectures in coverage and adaptability, while static and semi-static approaches remain optimal in compliance-critical, resource-constrained, and rigorously specified research domains.