W&D:Scaling Parallel Tool Calling for Efficient Deep Research Agents

Abstract: Deep research agents have emerged as powerful tools for automating complex intellectual tasks through multi-step reasoning and web-based information seeking. While recent efforts have successfully enhanced these agents by scaling depth through increasing the number of sequential thinking and tool calls, the potential of scaling width via parallel tool calling remains largely unexplored. In this work, we propose the Wide and Deep research agent, a framework designed to investigate the behavior and performance of agents when scaling not only depth but also width via parallel tool calling. Unlike existing approaches that rely on complex multi-agent orchestration to parallelize workloads, our method leverages intrinsic parallel tool calling to facilitate effective coordination within a single reasoning step. We demonstrate that scaling width significantly improves performance on deep research benchmarks while reducing the number of turns required to obtain correct answers. Furthermore, we analyze the factors driving these improvements through case studies and explore various tool call schedulers to optimize parallel tool calling strategy. Our findings suggest that optimizing the trade-off between width and depth is a critical pathway toward high-efficiency deep research agents. Notably, without context management or other tricks, we obtain 62.2% accuracy with GPT-5-Medium on BrowseComp, surpassing the original 54.9% reported by GPT-5-High.

• The paper introduces the W&D framework that enables parallel tool calling per reasoning step, reducing latency and token usage.
• It demonstrates significant efficiency gains with a 35.9% reduction in processing cost and improved accuracy across multiple benchmarks.
• Dynamic scheduling strategies, especially descending scheduler, enhance performance via redundant verification and effective query decomposition.

Scaling Parallel Tool Calling for Efficient Deep Research Agents: Analysis of the W&D Framework

Overview and Motivation

The paper "W&D: Scaling Parallel Tool Calling for Efficient Deep Research Agents" (2602.07359) introduces a novel approach for enhancing agentic systems tasked with complex, multi-step research and information-seeking workflows. While prior work predominantly focuses on scaling depth — increasing the number of sequential reasoning steps and tool calls — this study investigates the complementary dimension of width: issuing multiple tool calls simultaneously within a single reasoning step. The authors propose the Wide and Deep (W&D) agent framework, systematically evaluating the impact of parallel tool calling on both proprietary and open-source LLMs across challenging deep research benchmarks. The methodology is grounded in rigorous empirical analysis, including ablation studies on tool call schedulers and case studies exploring the roots of accuracy gains.

Figure 1: (a) Visual comparison of sequential (single) versus parallel tool calling traces in deep research agents. (b) Performance and iteration statistics showing consistent gains and efficiency from increasing the number of parallel tool calls per step.

Methodology: Parallel Tool Calling Formalization and Scheduling

The W&D agent extends traditional sequential agent traces by formalizing a parallel agent trace $\tau_{\text{par}}$, in which the LLM generates a set $\mathcal{A}^{\text{par}_t}$ of $m$ concurrent tool calls per step. These tool calls are executed in parallel, reducing wall-clock latency and token consumption due to fewer LLM reasoning steps per tool call.

To maintain precise control over tool call counts, the agent is prompted each iteration with a user message specifying the number of calls required, a strategy found to outperform static system instructions. The framework is fully compatible with modern proprietary LLMs (GPT-5, Gemini 3.0, Claude 4.5) and open-source models (DeepSeek-V3.2, Qwen3-235B), evaluated in the MCP-Universe agent environment with standardized tool access (search, scraping, Python interpreter).

Empirical Results: Accuracy and Efficiency Gains

Parallel tool calling — scaling width — yields statistically significant improvements in accuracy and efficiency across all tested benchmarks. On the BrowseComp dataset, parallel tool calling with three tools per turn yields 68% accuracy, reducing average turns required (23.8 vs 45.7 for single tool calling) and lowering both computation and API costs; for example, a 35.9% reduction in total processing cost and a 40.6% reduction in wall-clock time are achieved for 100 samples.

Figure 2: BrowseComp performance: accuracy versus average number of turns (left), and accuracy versus number of parallel tools per turn (right); parallel tool calling reduces turn count and increases accuracy.

Figure 3: Top row: GPT-5-medium performance scaling across multiple benchmarks with parallel tool calling. Bottom row: Cross-model analysis in BrowseComp validating parallel tool calling benefits.

Crucially, these benefits generalize to multiple benchmarks (BrowseComp, HLE, GAIA) and models, although the effect is less pronounced in open-source models, implying the need for further training optimization for agentic behaviors.

A contradictory finding is highlighted: GPT-5-Medium achieves 62.2% accuracy on the full BrowseComp set using parallel tool calling, outperforming the original GPT-5-High (54.9%) despite lacking context management or specialized tricks.

Analysis: Sources of Accuracy Gains in Parallel Tool Calling

Manual agent trace inspection identifies three key contributing factors to the performance improvement:

• Credible Information Aggregation: Parallel tool calling expands the search scope, allowing reasoning steps to compare sources and select authoritative information, mitigating erroneous reliance on a single API or database.
• Redundancy and Verification: Multiple tool calls serve as mutual verification points; if scraping or summarization tools return inconsistent data, the agent is able to detect unreliability and reinitiate searches, thus avoiding propagation of hallucinated or faulty data.
• Query Decomposition: The agent can break complex queries into simpler, parallelizable sub-queries, improving retrieval effectiveness for constraint-heavy or multi-faceted information requests.

Tool Call Scheduling: Balancing Exploration and Exploitation

Beyond constant-width parallel tool calling, the paper explores dynamic tool call scheduling strategies, denoting the number of calls per turn as $m_t$. Tested strategies include constant, ascending, descending, and automatic (LLM-decided) schedulers.

Key results:

• Descending scheduler (explore-then-exploit, more calls early, fewer later) achieves the highest accuracy (74% with 23.5 average turns), outperforming both constant and ascending strategies.
• The Automatic scheduler (LLM-decided) does not exceed descending, indicating current LLMs lack skill in autonomous optimal tool call balancing.

  Figure 4: Average number of tool calls per turn for different tool call schedulers, illustrating adherence to specified strategies and effect on agent behavior.

Practical and Theoretical Implications

The study demonstrates that scaling tool execution width via parallel tool calling is a critical axis for agentic optimization, often overlooked compared to context or reasoning depth scaling. The efficiency gains (reduced API cost, latency) are substantial, and the increased accuracy is rooted in redundancy, source verification, and fine-grained decomposition — mechanisms not inherent to the sequential agent design.

Practically, integrating parallel tool execution into agent frameworks is straightforward, requiring only prompting modifications rather than complex orchestration or sub-agent architecture. The demonstrated performance gains suggest parallel tool calling should be adopted as a default for high-throughput, accuracy-sensitive agentic use cases.

Theoretically, the inability of current LLMs to autonomously optimize the width-depth trade-off highlights a gap in agentic control learning, potentially addressable through reinforcement learning or dedicated scheduling objective functions. As proprietary and open-source LLMs evolve, training paradigms should include agentic tool call scheduling and reasoning coordination, particularly for multi-tool and multi-path environments.

Conclusion

The W&D agent framework offers a robust empirical and theoretical foundation for scaling parallel tool calling in deep research agents. Parallel execution not only enhances efficiency but also accuracy, driven by improved credibility verification, redundancy against tool failures, and query decomposition. Dynamic tool call scheduling, especially explore-then-exploit descent, further boosts agent effectiveness, though current LLMs require explicit prompting to leverage these strategies. Future research should focus on autonomous scheduling and reasoning coordination, laying the groundwork for highly efficient, adaptive agentic systems in complex research tasks.