PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Research Agents: A Systematic Examination And Roadmap

Yuxuan Huang, Yihang Chen, Haozheng Zhang, Kang Li, Huichi Zhou, Meng Fang, Linyi Yang, Xiaoguang Li, Lifeng Shang, Songcen Xu, Jianye Hao, Kun Shao, Jun Wang

arXiv:2506.18096v2cs.AI

TL;DR

Deep Research agents address the need for autonomous systems that can handle complex, multi-turn research through retrieval, reasoning, planning, tool use, and report generation. The survey analyses their technologies, workflows, optimisation methods, architectures, and benchmarks, and concludes that broader tool integration, parallel planning, and improved end-to-end optimisation remain central priorities.

  • Problem

    Existing research lacks a unified account of DR agents’ retrieval, reasoning, planning, tool-use, workflow, and evaluation dimensions.

  • Method

    The survey systematically reviews DR-agent technologies, architectures, optimisation approaches, industrial systems, and benchmarks, organising workflows and agent compositions into taxonomies.

  • Results

    The survey identifies static versus dynamic workflows, planning and agent-composition categories, and recurring limitations in current DR benchmarks and execution designs.

  • Takeaways & Limitations

    Future DR-agent development should prioritise flexible tool integration, asynchronous and parallel planning, multimodal evaluation, and end-to-end optimisation for multi-agent systems.

  • Takeaways & Limitations

    Current DR evaluations often rely on static-corpus QA or narrow task metrics, under-assessing open-web research and structured multimodal report generation.

Abstract

from arXiv · show

The rapid progress of Large Language Models (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}.

1 Introduction

Deep Research agents extend LLMs with dynamic reasoning, adaptive planning, iterative tool use, and real-time external information access for complex research tasks. The survey reviews their architectures, technologies, benchmarks, and open challenges.

  • Deep Research agents: Deep Research agents combine LLM reasoning, dynamic planning, adaptive web interaction, and analytical tools to perform autonomous research.They retrieve, aggregate, and analyse external information for open-ended informational tasks.
  • Survey scope: The survey analyses search integration, tool invocation, workflow architectures, optimisation methods, and representative industrial implementations.Its scope includes retrieval mechanisms, system architectures, tool use, performance characteristics, and tuning paradigms.
  • Taxonomy: A unified taxonomy classifies DR systems by static versus dynamic workflows, planning strategies, and single-agent versus multi-agent architectures.The framework is intended to bridge diverse technical methodologies and industrial solutions.
  • Benchmarks: The survey reviews benchmarks that assess retrieval accuracy, reasoning depth, and adaptive tool invocation in DR systems.It systematically categorises existing evaluation approaches and their coverage of critical capabilities.
  • Open challenges: Open challenges include expanding retrieval scope, enabling asynchronous parallel execution, developing multimodal benchmarks, and improving multi-agent robustness and efficiency.These directions are presented as priorities for future DR-agent research.

2 Background and Preliminaries

The background positions Deep Research agents as an evolution of reasoning-enhanced retrieval systems, while MCP and A2A address interoperability for tools and collaborating agents. Agentic retrieval improves adaptation but retains computational, interpretability, and corpus-dependence limitations.

  • Reasoning models: Large reasoning models improve complex problem solving through architectural and training advances such as Chain-of-Thought prompting.Related frameworks also support specialised external knowledge and computational resources within reasoning workflows.
  • Retrieval-augmented generation: Traditional RAG retrieves external documents through relatively static pipelines before generating answers from the retrieved passages.Its primary role is mitigating hallucinations and improving information-search accuracy.
  • Agentic retrieval: Agentic RAG adds iterative retrieval, adaptive querying, and dynamic workflow adjustment to improve multi-step reasoning.Examples include reinforcement-learning query refinement and graph-based multi-hop retrieval.
  • Agentic retrieval: Agentic RAG remains limited by computational overhead, user-intent alignment, interpretability, and reliance on pre-existing or periodically updated corpora.These constraints affect its ability to adapt across complex research requirements.
  • MCP and A2A: MCP standardises external-tool access, while A2A orchestrates collaboration among agents, together forming a modular and scalable interoperability foundation.The separation addresses inconsistent APIs, maintenance costs, and redundant development across tool-use systems.

3 Deep Research: Search Engine, Tool Use, Workflow, Tuning, Non-parametric Continual Learning

This section examines DR agents through retrieval, tool use, workflow, and optimisation components. It contrasts fixed RAG pipelines with DR systems that dynamically retrieve information, invoke tools, and adapt multi-stage research execution.

  • Search and retrieval: DR agents extend conventional RAG by combining dynamic retrieval, real-time tool use, and adaptive reasoning in one system.This design supports greater autonomy and context-awareness for complex, multi-step queries and changing contexts.
  • Search and retrieval: The survey compares API-based search interfaces with browser-based exploration as alternative mechanisms for dynamic knowledge acquisition.The comparison is presented as part of the search-engine integration component of DR-agent design.
  • Tool use: Tool-use capabilities encompass code execution, mathematical computation, file manipulation, and multimodal processing within the agent inference pipeline.These modules are treated as core components for developing and optimising DR agents.
  • Workflow architecture: Workflow analysis covers static and dynamic designs, planning strategies, single-agent and multi-agent organisation, memory mechanisms, and auxiliary components.Planning strategies include planning-only, intent-to-planning, and unified intent-planning interaction types.

3.1 Search Engine: API vs. Browser

DR agents acquire external knowledge through API-based or browser-based search, with APIs emphasizing efficient scalable retrieval and browsers enabling interactive access to dynamic, multimodal, and unstructured content.

  • Comparative Overview: The survey compares DR agents across search engines, base models, release dates, and benchmark categories including GAIA, HLE, and other question-answering tasks.Table 1 organizes whether systems use API-based or browser-based search alongside their models and evaluation benchmarks.
  • API-Based Search: API-based search engines provide fast, efficient, scalable retrieval with relatively low latency and computational overhead.Examples include coordinated Google Search and arXiv APIs, continuously updated indexes, and open-source DuckDuckGo pipelines.
  • Adaptive Retrieval: Existing systems combine search interfaces with learned decisions about when to search, what to search, and how intensely to retrieve evidence.Reported approaches include query decomposition, evidence integration during reasoning, and reinforcement learning that adapts search intensity to task difficulty.
  • Browser-Based Search: Browser-based search simulates human interaction with web pages to retrieve dynamic, flexible, multimodal, and unstructured content.Browser agents may open tabs, issue queries, click links, scroll, fill forms, execute JavaScript, and download files for analysis.

3.2 Tool Use: Empowering Agents with Extended Functionalities

DR agents extend their inference pipelines with code execution, analytics, multimodal processing, computer use, and interoperable tool frameworks to handle varied research tasks.

  • Tool Modules: Tool-use evaluations compare DR agents across capabilities such as code interpretation, data analytics, multimodal processing, and related integrations.The survey identifies these modules as central mechanisms for interacting with external environments and data sources.
  • Code Interpretation: Code interpreters let agents execute scripts during inference for data processing, algorithm verification, model simulation, and computational reasoning.Most surveyed DR agents embed script-execution environments, commonly using Python or Java utilities.
  • Data Analytics: Data analytics modules convert retrieved material into structured insights through summary statistics, visualizations, table generation, and quantitative evaluation.These capabilities support hypothesis testing and decision-making, locally or through remote services.
  • Multimodal Processing: Multimodal tools integrate and generate text, images, audio, and video within a unified reasoning pipeline, but adoption remains concentrated in a subset of mature systems.Most academic prototypes have not implemented multimodal processing and generation.
  • Computer Use: Computer-use systems expand DR agents toward autonomous interaction with web environments, code, external APIs, and iterative self-reflection.AutoGLM Rumination is described as an RL-based example combining self-reflection, refinement, and advanced function calling.

3.3 Architecture and Workflow

DR architectures are organized by workflow flexibility, planning strategy, agent composition, and memory management, contrasting predefined pipelines with adaptive systems that reconfigure tasks as contexts change.

  • Static Workflows: Static workflows execute manually predefined sequential subtasks and suit well-defined, structured research scenarios.Examples include separate ideation, experimentation, and reporting phases in AI Scientist.
  • Static Workflows: Static workflows have limited generalization because each distinct task requires a specifically tailored pipeline.Their structured clarity and ease of implementation come with reduced flexibility across varied tasks.
  • Dynamic Workflows: Dynamic workflows adapt task structures through automated planning, iterative refinement, interactive allocation, and replanning as feedback or context evolves.They are presented as more general and adaptable than static workflows.
  • Planning Strategies: Planning strategies include planning-only, intent-to-planning, and unified intent-planning, differing in whether clarification and user confirmation occur before execution.Planning-only uses the initial prompt; intent-to-planning clarifies intent first; unified intent-planning also requests plan confirmation.
  • Agent Composition: Dynamic systems may use a single agent that integrates planning and execution or multiple specialized agents coordinated through adaptive task allocation.Multi-agent systems commonly use hierarchical or centralized planners, while single-agent systems maintain a unified cognitive loop.
  • Memory Mechanisms: Memory mechanisms reduce redundant queries and improve task efficiency and coherence by retaining information across retrieval rounds.The survey discusses context expansion, intermediate-step compression, and external structured storage as optimization strategies.
  • Memory Mechanisms: Extending context windows is straightforward but can impose high computational costs and inefficient resource utilization.Gemini is cited as supporting a context window of up to one million tokens with RAG.
  • Memory Mechanisms: Intermediate compression and external storage reduce token pressure by summarizing reasoning or preserving historical results beyond the context window.Examples include summarized workflow outputs, document compression, and external file systems for intermediate outcomes.

3.4 Tuning: Beyond Prompting toward Capability Enhancement

This section surveys tuning strategies that move DR agents beyond prompting, covering supervised fine-tuning, reinforcement learning, and emerging self-evolution approaches. It compares methods, training signals, base models, datasets, and implementation patterns across representative systems.

  • 3.4 Tuning: Beyond Prompting toward Capability Enhancement: Table 3 compares DR agents by supervised or reinforcement-learning method, base model, training data, and reward design.The listed systems include SimpleDeepSearcher, ReSearch, Search-R1, WebShaper, WebWatcher, and others.
  • 3.4 Tuning: Beyond Prompting toward Capability Enhancement: Prompt-based methods enable complex functions without costly fine-tuning, but their performance remains bounded by the backbone LLM and difficult-to-optimise workflows.Increasing decision-process complexity can quickly reach the model’s performance ceiling.
  • 3.4.1 SFT-based Optimization: SFT methods target search-query formulation, structured report generation, dynamic retrieval planning, information synthesis, and integrated tool use.Open-RAG adds retrieval, relevance, grounding, and utility tokens, while later methods use rejection sampling or selected expert steps.
  • 3.4.1 SFT-based Optimization: SFT improves generalisation but remains confined to offline, static retrieval pipelines characteristic of retrieval-augmented systems.This limitation motivates reinforcement learning for online query generation and tool invocation.
  • 3.4.2 Reinforcement Learning-based Optimisation: RL-based methods strengthen iterative retrieval, dynamic tool invocation, and integrated reasoning by learning from real-time rewards.The surveyed progression runs from query optimisation through adaptive reasoning and search-reasoning integration toward end-to-end agent training.
  • 3.4.2 Reinforcement Learning-based Optimisation: Three RL implementation patterns are identified: proprietary industrial systems, modular academic optimisation, and hybrid multi-task systems.SimpleDeepSearcher combines process-based rewards with training across 6 QA datasets, while Qwen2.5 and LLaMA3 are prevalent base-model families.
  • 3.4.2 Reinforcement Learning-based Optimisation: Rule-based rewards commonly encode retrieval relevance, information accuracy, or successful tool invocation, with PPO and GRPO used for policy optimisation.These reward and optimisation choices expose task objectives explicitly in many open-source implementations.
  • 3.4.2 Reinforcement Learning-based Optimisation: 12 to 3 gradient direction conflicts per training epoch are reported for GRPO, alongside faster convergence and broader reward-distribution coverage than PPO.The passage also reports faster KL-divergence stabilisation during alignment.

3.5 Non-parametric Continual Learning

Non-parametric continual learning offers runtime adaptation without updating model parameters, with case-based reasoning as a prominent trajectory-level approach for DR agents.

  • 3.5 Non-parametric Continual Learning: Parameter-based SFT and RL face challenges from model scaling, structured experience requirements, and increasingly complex training algorithms.These constraints motivate runtime alternatives for agents with hierarchical workflows.
  • 3.5 Non-parametric Continual Learning: Non-parametric continual learning refines agent capabilities at runtime with minimal data and computational overhead.This makes it suitable for DR agents with complex architectures.
  • 3.5 Non-parametric Continual Learning: Case-based reasoning retrieves, adapts, and reuses structured problem-solving trajectories from an external case bank.Unlike static RAG databases, CBR supports online contextual adaptation and task-level generalisation; Agent K adds reward-guided memory reuse.
  • 3.5 Non-parametric Continual Learning: Compared with fixed prompts and static RAG corpora, CBR performs dynamic trajectory retrieval with reasoning-centred memory organisation.The survey presents these self-evolution paradigms as promising for knowledge reuse and continual learning despite limited current attention.

4 Industrial Applications of Deep Research Agents

Industrial DR agents increasingly combine adaptive planning, multimodal retrieval, tool integration, and structured reporting across single-agent and modular systems. The surveyed examples illustrate different ways to automate complex research and analysis workflows.

  • 4.1 Open AI Deep Research: OpenAI DR uses a single-agent, reinforcement-learning-fine-tuned o3 model to clarify intent, plan multi-step research, browse, retrieve multimodal information, and compute.Its reported advances include iterative workflow adaptation, context memory, multimodal processing, and integrated browsing and programming tools.
  • 4.3 Perplexity Deep Research: Perplexity’s DR agent decomposes complex queries, performs iterative targeted searches, evaluates authoritative sources, and synthesises structured reports.Its workflow emphasises repeated retrieval with dynamic adjustment.
  • 4.4 Grok DeepSearch: Grok DeepSearch combines real-time retrieval with multimodal reasoning, credibility filtering, concurrent cross-source verification, and structured outputs with citations.It can alternate adaptively between lightweight retrieval and intensive analysis modes.
  • 4.5 Microsoft Researcher and Analyst: Microsoft’s Researcher and Analyst agents access work data and web information, supporting multi-step research, market strategy, reporting, and professional data analytics.Researcher integrates internal and external data, while Analyst transforms raw data into insights using step-by-step reasoning.
  • 4.7 Kimi K2 Deep Research: Kimi K2 combines token-efficient learning, targeted data engineering, sparse architecture, and tool-aligned post-training for deep research.Its post-training uses supervised tool-use trajectories with sandboxed code execution and reinforcement learning with verifiable or rubric-based rewards.

5 Benchmarks for DR Agent

DR-agent benchmarks span factual question answering, long-horizon task execution, research and coding, and emerging GUI interaction. However, comprehensive evaluation must capture the full multi-step workflow and structured evidence-grounded reporting expected of DR agents.

  • Benchmark Scope: DR evaluation should measure multi-step retrieval, cross-source synthesis, dynamic tool invocation, and structured evidence-grounded report generation.These capabilities define the full research workflow rather than isolated answer production.
  • Benchmark Gaps: Long-form datasets such as Qasper and ELI5 test extended coherence but do not align with structured evidence-based reporting expected of DR agents.The survey therefore identifies a need for specialized benchmarks covering the multimodal, multi-stage nature of DR workflows.
  • QA Benchmarks: QA benchmarks range from single-hop factual recall to multi-hop and research-style question answering.SimpleQA, TriviaQA, and PopQA emphasize factual recall, while NQ and TELEQnA add document and domain complexity.
  • Task Execution Benchmarks: Task-execution benchmarks assess integrated tool use, environment perception, information filtering, and end-to-end workflow completion.General-purpose tasks include GAIA, AssistantBench, and Magentic-One; research-oriented tasks include software, machine-learning, and scientific-experiment benchmarks.
  • GUI-Based Evaluation: GUI environments such as OSWorld, WebArena, and SpaBench extend evaluation toward direct manipulation of applications and web interfaces.These environments support testing embodied research capabilities in realistic user-facing settings.

6 Challenge and Future Directions

The roadmap focuses on expanding information access, improving browser and execution efficiency, strengthening tool-integrated reasoning, and developing better benchmark alignment. It also highlights multi-agent coordination and self-evolving workflows as open research directions.

  • Broaden Information Source: Current DR agents are constrained by static repositories or conventional search interfaces and often cannot access proprietary applications, specialized databases, or subscription-only services.The survey proposes broader modular tool integration through MCPs to reach resources beyond standard browsers and search engines.
  • Broaden Information Source: After proprietary data access improves, human-centered browsers remain a bottleneck because visual rendering, fragile locators, and anti-bot defenses hinder automated interaction.This shifts the workflow’s rate-limiting factor from data acquisition toward webpage interaction efficiency.
  • Asynchronous Parallel Execution: Most DR agents execute subtasks linearly, motivating asynchronous parallel architectures that exploit parallelism and support dynamic execution adjustments.The proposed direction targets efficiency and robustness constraints associated with purely sequential planning.
  • Tool-Integrated Reasoning: Tool-integrated reasoning requires agents to adapt reasoning pathways to intermediate results, while supervised fine-tuning can cause overreasoning or inappropriate tool selection.Reinforcement learning is discussed as a route toward more adaptive tool-based reasoning.
  • Benchmark Misalignment: Existing evaluations can reward answering from memorized static-corpus knowledge, so open-web, time-sensitive, and continually refreshed benchmarks are needed.BrowseComp is identified as a step toward forcing retrieval of difficult-to-find online information.
  • Benchmark Misalignment: Current metrics under-assess structured multimodal reports that combine narrative, tables, figures, and citations, motivating end-to-end report-generation benchmarks.Such benchmarks should assess factual accuracy, discourse structure, and cross-modal alignment together.
  • Parametric Optimisation of Multi-Agent Architectures: Single-agent systems place planning, tool invocation, and report generation on one backbone model, while multi-agent systems introduce unresolved training and coordination challenges.Hierarchical reinforcement learning and staged post-training are proposed as possible optimization directions.
  • Self-Evolving Language Model Agents: Self-evolution remains underdeveloped, with current efforts focused narrowly on case-based reasoning; future work should expand experience reuse and mutable workflow optimization.Suggested representations include hierarchical experience traces and workflow trees or graphs.

7 Conclusion

The survey synthesizes Deep Research agents as an emerging paradigm for automated research support and organizes methods by optimization approach, information retrieval, and report generation. It also reviews representative industrial systems and their strengths and limitations.

  • Conclusion: The survey reviews DR agents through prompt-based, fine-tuning-based, and reinforcement-learning-based approaches for information retrieval and report generation.It also examines representative industry systems, including their technical implementations, strengths, and limitations.
  • Conclusion: Prompt-based methods support efficient, cost-effective deployment, whereas fine-tuning and reinforcement learning optimize model parameters for reasoning and decision-making.The contrast is presented as a distinction between deployment efficiency and explicit parameter optimization.

Limitation

Despite progress, DR agents still face limited generalization, inflexible workflows, difficult granular-tool integration, and substantial computational complexity in planning and optimization.

  • Limitation: Key limitations include limited generalization across diverse tasks, inflexible workflows, difficult integration of granular external tools, and substantial computational complexity.The conclusion connects these limitations to proposed directions involving modular capability providers and asynchronous or parallel planning.
Loading 2506.18096v2…