Agentic Deep-Research System

• Agentic Deep-Research Systems are hierarchical, modular architectures that enable LLMs and agents to autonomously perform complex, multi-step research tasks.
• They combine planning, specialized sub-agents, and dynamic context management to overcome limitations of static designs and achieve state-of-the-art benchmarks.
• The architecture features a proactive supervisor module with anomaly detection and self-correction protocols to ensure robust and scalable research execution.

Agentic Deep-Research System

Agentic deep-research systems are hierarchical, modular architectures designed to enable LLMs and associated agents to autonomously perform complex, multi-step research tasks. These systems tightly integrate planning, tool-use, sub-agent collaboration, dynamic context management, and real-time supervision, overcoming the limitations of classical retrieval-augmented generation (RAG) and static pipeline designs for open-ended, long-horizon problems. By orchestrating basic tools and specialized sub-agents within an extensible framework, agentic deep-research systems achieve state-of-the-art performance across web, enterprise, and scientific research benchmarks (Cai et al., 27 Jan 2026).

1. System Architecture and Hierarchical Design

Agentic deep-research systems, exemplified by Yunque DeepResearch, employ a three-tiered architecture:

1. Main Agent & Multi-Agent Orchestration System (MAOS): Central controller for global intent analysis, planning, and orchestration. Decomposes the input research query into explicit sub-goals and dispatches them to appropriate capabilities.
2. Atomic Capability Pool (ACP): Collection of both Basic Tools (e.g., search, file parsing, code execution) and Specialized Sub-Agents (e.g., browser-interaction agents, data analysis agents). Each atomic task (such as a web search) is handled by a basic tool, while more complex tasks are delegated to sub-agents with specialized policies.
3. Supervisor Module: Proactive monitor for anomaly detection, error diagnosis, and dynamic recovery. Functions independently of the orchestration pipeline, ensuring resilience via online self-correction and context pruning.

Data and Control Flow (cf. (Cai et al., 27 Jan 2026)):

• The user issues a research query $Q$ to the Main Agent, which decomposes it into sub-goals $\{g_1, \dots, g_n\}$. For each $g_i$, the routing mechanism ($P(r \mid g, C)$) probabilistically selects the optimal resource $r^*$ based on a cost model over tools/sub-agents, balancing latency and expected error.
• Upon execution, tool/sub-agent responses $o_i$ are incorporated into the evolving context $C_t$ via the Dynamic Context Manager.
• The Supervisor computes an anomaly score $S_i$ per step, triggering context pruning and plan regeneration if $S_i > \tau_s$.
• After all sub-goals, the Main Agent synthesizes a final answer $A$.

2. Multi-Agent Orchestration and Atomic Capability Pool

The multi-agent orchestration system routes each sub-goal by evaluating whether it is atomic or requires a specialized sub-agent.

Routing Model:

$\{g_1, \dots, g_n\}$0

where $\{g_1, \dots, g_n\}$1 (the pool of tools and sub-agents), and $\{g_1, \dots, g_n\}$2 can be a composite of expected latency, error likelihood, or resource utilization.

Registration and Extensibility:

• New tools or sub-agents are registered via a standardized interface (can_handle, execute methods), with metadata and optional cost functions specified in a JSON registry.
• ACP can be extended dynamically. For instance, adding an “ExcelAnalysisAgent” or “PDFSummarizerTool” requires interface implementation and registry entry; the system auto-discovers new capabilities at runtime (Cai et al., 27 Jan 2026).

3. Dynamic Context Management

Agentic deep-research must efficiently manage long-horizon context to avoid explosion of irrelevant intermediate data.

Memory and Summarization

Each completed sub-goal $\{g_1, \dots, g_n\}$3 is encapsulated as a memory unit: $\{g_1, \dots, g_n\}$4

• $\{g_1, \dots, g_n\}$5: interaction rounds
• $\{g_1, \dots, g_n\}$6: tool-use logs
• $\{g_1, \dots, g_n\}$7: semantic summary via a summarization function $\{g_1, \dots, g_n\}$8

Pruning and Complexity Reduction

The context window $\{g_1, \dots, g_n\}$9 is managed by folding completed sub-goals into compact semantic summaries and dropping raw traces when their total token length exceeds a threshold $g_i$0. This reduces memory growth from $g_i$1 rounds to $g_i$2 memory units, with $g_i$3 (Cai et al., 27 Jan 2026).

Custom summarizers may be plugged in by subclassing and configuring the Context Manager.

4. Proactive Supervision and Robustness

The Supervisor Module provides continual anomaly detection and active recovery at each execution step.

Anomaly Score

$g_i$4

• RepetitionScore: Normalized count of repetitive actions or chain-of-thought traces.
• SyntaxErrorCount: Rate of malformed tool calls.
• LowUtilityRatio: Fraction of unsuccessful tool invocations over a sliding window.

If $g_i$5, the Supervisor interrupts normal execution and initiates the self-correction protocol.

Self-Correction Protocol

1. Diagnosis: Agent is prompted to analyze the last $g_i$6 steps and identify faults.
2. Trajectory Pruning: Last $g_i$7 context turns removed.
3. Regeneration: Main Agent re-plans with the pruned context, producing a revised action sequence.

This loop minimizes cascading errors and enhances resilience in long, multi-hop research trajectories (Cai et al., 27 Jan 2026).

5. Evaluation and Benchmark Performance

Yunque DeepResearch achieves state-of-the-art Pass@1 accuracy across multiple research benchmarks:

Method	GAIA	BrowseComp	BrowseComp-ZH	HLE
Gemini-3-Pro + ReAct	73.8	52.5	67.1	45.8
WebResearcher (open-source)	72.8	37.3	45.2	28.8
OpenAI DeepResearch	67.4	51.5	42.9	26.6
Yunque DeepResearch	78.6	62.5	75.9	51.7

Ablation Results (GAIA, BrowseComp, BrowseComp-ZH, HLE):

• Removal of the memory module significantly degrades performance on BrowseComp (-10.4) and BrowseComp-ZH (-7.4).
• Disabling the Supervisor reduces robustness across all tasks, especially GAIA (-8.7) and BrowseComp-ZH (-10.5).
• Removing the Browser-Use GUI Agent or Data Analysis Agent impacts specific modalities.

A time cap of 1.5 hours per question and a maximum of 75 tool calls are enforced during evaluation (Cai et al., 27 Jan 2026).

6. Implementation, APIs, and Extensibility

Yunque DeepResearch leverages:

• LLM Backbone: Gemini-3-Pro via Tencent LLM-API.
• Tool Framework: Python Flask microservices for basic tools (search, parsing, code execution in Docker sandboxes).
• Browser-Use Agent: Puppeteer with Headless Chrome, LLM-guided policies.
• Data Analysis Agent: Python/Pandas/Numpy in a secure container.
• Supervisor & ContextManager: Modular Python components under the orchestration of the Main Agent.

Open-source code (Apache 2.0) and implementations are available for community extension.

Extending Capabilities: Register new tools/sub-agents by implementing designated interfaces and updating ACP registries and cost functions. The orchestration server supports automatic capability discovery.

Extending Context Management: Plug-in custom summarizers by subclassing and configuration.

7. Significance and Paradigm Positioning

Agentic deep-research systems, as represented by Yunque DeepResearch, establish a new paradigm for autonomous research agents:

• Hierarchical orchestration and context-aware memory radically improve multi-step, open-ended task performance in LLMs.
• Explicit feedback control and dynamic pruning guard against error propagation, a central challenge in long-horizon automation.
• Modularity and standardized interfaces enable rapid extension for new domains and tools.
• Open-source release and empirical benchmarks enhance reproducibility, transparency, and community-driven advancement.

These architectures encode the prevailing best practices for robust, scalable, and extensible autonomous research agents in real-world information-heavy domains (Cai et al., 27 Jan 2026).