Deep Research Agents: A Systematic Examination And Roadmap

Abstract: The rapid progress of LLMs has given rise to a new category of autonomous AI systems, referred to as Deep Research (DR) agents. These agents are designed to tackle complex, multi-turn informational research tasks by leveraging a combination of dynamic reasoning, adaptive long-horizon planning, multi-hop information retrieval, iterative tool use, and the generation of structured analytical reports. In this paper, we conduct a detailed analysis of the foundational technologies and architectural components that constitute Deep Research agents. We begin by reviewing information acquisition strategies, contrasting API-based retrieval methods with browser-based exploration. We then examine modular tool-use frameworks, including code execution, multimodal input processing, and the integration of Model Context Protocols (MCPs) to support extensibility and ecosystem development. To systematize existing approaches, we propose a taxonomy that differentiates between static and dynamic workflows, and we classify agent architectures based on planning strategies and agent composition, including single-agent and multi-agent configurations. We also provide a critical evaluation of current benchmarks, highlighting key limitations such as restricted access to external knowledge, sequential execution inefficiencies, and misalignment between evaluation metrics and the practical objectives of DR agents. Finally, we outline open challenges and promising directions for future research. A curated and continuously updated repository of DR agent research is available at: {https://github.com/ai-agents-2030/awesome-deep-research-agent}.

• The paper surveys Deep Research agents, detailing their integration of LLM reasoning with dynamic retrieval and multi-agent protocols.
• It systematically compares DR agent architectures, highlighting trade-offs between static and dynamic workflows and different planning strategies.
• The study outlines future research directions, emphasizing challenges in retrieval scope, multimodal benchmarks, and agent interoperability.

Deep Research Agents: A Systematic Examination And Roadmap

This paper (2506.18096) presents a comprehensive survey and analysis of Deep Research (DR) agents, which are AI systems powered by LLMs and designed to tackle complex informational research tasks. The survey covers the foundational technologies, architectural components, benchmarks, and future research directions in the rapidly evolving field of DR agents.

Background and Foundational Concepts

The paper begins by reviewing the advances in reasoning capabilities of LLMs, particularly focusing on Chain-of-Thought (CoT) prompting and its variants. It highlights the limitations of current reasoning frameworks, such as hallucinations and static knowledge, which motivate the need for DR agents that integrate external information sources and real-time retrieval mechanisms. The survey also discusses the evolution of Retrieval-Augmented Generation (RAG) from static pipelines to agentic RAG, which incorporates iterative retrieval and dynamic workflow adjustments. It also emphasizes the importance of Model Context Protocol (MCP) and Agent-to-Agent (A2A) protocols for enabling interoperability and collaboration in multi-agent systems.

Core Components and Methodologies

The survey explores the core components that constitute DR agents (Figure 1):

Figure 1: A structural overview of a DR agent in an multi-agent architecture for ease of illustration.

It compares API-based and browser-based search engine integration, highlighting the trade-offs between efficiency and comprehensiveness. The paper also examines tool use capabilities, such as code interpreters, data analytics, and multimodal processing, which enable DR agents to interact with external environments and generate structured insights. The classification of DR agent architectures based on static versus dynamic workflows, planning strategies, and single-agent versus multi-agent configurations is systematically analyzed (Figure 2).

Figure 2: Comparison of DR Workflows: (1) Static vs. Dynamic Workflows: Static workflows rely on predefined task sequences, while dynamic workflows allow LLM-based task planning. (2) Planning Strategies: Three types include: planning-only (direct planning without clarifying user intent), intent-to-planning (clarifying intent before planning), and unified intent-planning (generating a plan and requesting user confirmation). (3) Single-Agent vs. Multi-Agent: Dynamic workflows can be categorized to dynamic-multi-agent systems (tasks distributed across specialized agents) or dynamic-single-agent systems (a LRM autonomously updates and executes tasks).

The memory mechanisms employed by DR agents for long-context optimization, including context window expansion, intermediate step compression, and external structured storage, are explored. Furthermore, the paper discusses tuning methodologies such as prompt-driven structured generation, LLM-driven prompting, fine-tuning strategies, and reinforcement learning approaches aimed at optimizing agent performance. The survey also touches on non-parametric continual learning techniques that allow DR agents to self-evolve by dynamically adapting external tools and workflows.

Industrial Applications and Benchmarks

The survey provides an overview of several industrial applications of DR agents, including OpenAI Deep Research, Gemini Deep Research, Perplexity Deep Research, Grok DeepSearch, and Microsoft Copilot Researcher and Analyst. It highlights the key advancements and technological implementations of these systems. The benchmarks used for evaluating DR systems are categorized into question-answering and task execution scenarios, with a critical evaluation of their limitations and misalignment with the practical objectives of DR agents.

Challenges and Future Directions

The paper identifies several open challenges and promising directions for future research (Figure 3):

Figure 3: An overview of DR agents evolution over years.

These include expanding retrieval scope beyond traditional methods, enabling asynchronous parallel execution, developing comprehensive multi-modal benchmarks, and optimizing multi-agent architectures for enhanced robustness and efficiency. The importance of addressing these challenges to enable DR agents to function as truly autonomous and adaptable research assistants is emphasized.

Conclusion

The survey concludes by summarizing the advancements in DR agents and highlighting their potential to transform complex research workflows and drive innovation across various domains. It emphasizes the need for continued research to address the identified challenges and unlock the full potential of DR agents.