deep_search_r.json

Overview

deep_search_r.json defines a comprehensive Multi-Agent Deep Research Agent configuration designed for executing structured, multi-step research investigations. This agent is specialized for professional domains such as sales, marketing, policy, or consulting. Its core functionality is to orchestrate multiple specialized sub-agents to collaboratively:

  1. Discover high-quality, authoritative URLs (sources).

  2. Extract and structure detailed content from these URLs.

  3. Synthesize the extracted content into a polished, strategic research report resembling consulting firm deliverables (e.g., McKinsey-style reports).

The file encodes a detailed agent workflow, component parameters, failure recovery strategies, quality gates, and example processes, supporting multilingual capabilities and adaptive resource allocation.

Detailed Explanation of Components

The file uses a JSON format describing a DSL (Domain Specific Language) specification of the multi-agent pipeline, including the following key components:

Top-Level Fields

Key Components in dsl.components

  1. Agent:NewPumasLick — The Lead Agent ("Deep Research Agent")

    • Role: Strategy Research Director coordinating the entire multi-agent research.

    • Parameters:

      • llm_id: Uses "qwen-max@Tongyi-Qianwen" model for reasoning.

      • max_tokens: 4096 tokens per round.

      • max_retries: 3 attempts per subtask.

      • max_rounds: Up to 3 rounds of interaction.

      • presence_penalty and frequency_penalty enabled with values 0.5.

      • prompts: Takes the user query as input with role user.

      • sys_prompt: Extensive detailed prompt specifying mission, execution framework, research process, query classification, quality gates, adaptive strategies, language adaptation, and execution examples.

    • Downstream components: Sends output to Message:OrangeYearsShine.

    • Upstream: Receives from begin.

    • Functionality: Orchestrates the research process across three stages:

      • URL discovery via Web Search Specialist.

      • Content extraction via Content Deep Reader.

      • Strategic report generation via Research Synthesizer.

    • Usage example:

      {
  "sys.query": "How will AI transform healthcare by 2030?"
}

      The Lead Agent breaks down this query, classifies it, allocates tasks, manages sub-agents, and synthesizes final report content.

  2. Agent:FreeDucksObey — Web Search Specialist

    • Role: Expert in discovering high-quality URLs.

    • Parameters:

      • Uses "qwen-plus@Tongyi-Qianwen" model.

      • Executes only one search round (max_rounds=1).

      • Uses web search tools (e.g., TavilySearch) exclusively.

      • Does not extract content — strictly URL discovery.

      • Enforces quality criteria avoiding paywalled, low-authority, or outdated sources.

      • Outputs exactly 5 URLs with extraction guidance.

      • Includes detailed system prompt specifying mission, process, search strategy, quality criteria, output format, and collaboration protocols.

    • Downstream: Passes URLs to Content Deep Reader.

    • Usage example:

      • For query "AI regulation finance 2024", it performs multiple web searches, evaluates URLs, and outputs 5 prioritized links with extraction focus.

  3. Agent:WeakBoatsServe — Content Deep Reader

    • Role: Extracts structured and comprehensive content from the URLs provided by the Web Search Specialist.

    • Parameters:

      • Uses "moonshot-v1-auto@Moonshot" LLM.

      • Allows up to 3 rounds per URL extraction.

      • Uses web extracting tools (e.g., TavilyExtract).

      • Extracts full text, key statistics, findings, expert quotes, supporting data, methodology, and credibility scores.

      • Includes fallback strategies for extraction failures.

      • Outputs standardized EXTRACTED_CONTENT for each URL.

      • System prompt defines mission, process, extraction schema, quality assessment, failure handling, and collaboration.

    • Downstream: Passes structured content to Research Synthesizer.

    • Usage example:

      EXTRACTED_CONTENT:
URL: https://www.sec.gov/ai-guidance-2024
TITLE: "SEC Guidance on AI in Financial Services - March 2024"
FULL_TEXT: "The Securities and Exchange Commission (SEC) today announced comprehensive guidance..."
KEY_STATISTICS: 65% of banks now use AI, $2.3B investment in 2024
MAIN_FINDINGS: New compliance framework requires AI audit trails...
...

  4. Agent:SwiftToysTell — Research Synthesizer

    • Role: Final integration specialist that synthesizes extracted content into executive-grade, consulting-style strategic reports.

    • Parameters:

      • Uses "moonshot-v1-128k@Moonshot" model for deep analysis.

      • Supports multiple rounds to refine analysis.

      • Receives ANALYSIS_INSTRUCTIONS from Lead Agent.

      • Produces reports adhering strictly to analysis framework (e.g., McKinsey-style or BCG-style).

      • Enforces output constraints: no raw data, no intermediate summaries, only polished final reports.

      • System prompt includes mission, process, integration strategy, report structure templates, quality standards, synthesis techniques, and output formatting.

    • Downstream: Final output sent to Message.

    • Usage example:

      • Converts extracted content into sections: Executive Summary, Analysis, Recommendations.

      • Provides actionable insights, risk assessment, and implementation roadmap.

  5. Message:OrangeYearsShine

    • Role: Outputs the final content generated by Agent:NewPumasLick to the user interface or downstream system.

    • Parameters: Contains the content from the Lead Agent.

    • Upstream: Connected to Agent:NewPumasLick.

  6. Begin

    • Role: Entry point into the pipeline.

    • Downstream: Starts the Lead Agent component.

Tools Used

Global Variables

Important Implementation Details and Algorithms

Interaction with Other Parts of the System

Usage Example (End-to-End)

  1. User inputs query: "What are the main factors driving cryptocurrency market volatility?"

  2. Lead Agent analyzes the query, classifies it as Depth-first, and develops a plan involving multiple perspectives.

  3. Web Search Specialist discovers 5 high-quality URLs covering technical, regulatory, and psychological aspects.

  4. Content Deep Reader extracts full content from these URLs with structured metadata.

  5. Research Synthesizer integrates the content into a ~2000-word McKinsey-style strategic report with actionable insights.

  6. Final report is delivered through the Message component to the user.

Visual Diagram

flowchart TD
    Begin --> LeadAgent[Agent:NewPumasLick<br/>Lead Agent]
    LeadAgent -->|Dispatch URLs| WebSearch[Agent:FreeDucksObey<br/>Web Search Specialist]
    LeadAgent -->|Dispatch URLs| ContentReader[Agent:WeakBoatsServe<br/>Content Deep Reader]
    LeadAgent -->|Receive Extracted Content| ResearchSynth[Agent:SwiftToysTell<br/>Research Synthesizer]
    LeadAgent -->|Send Final Report| Message[Message:OrangeYearsShine]
    
    WebSearch -->|URLs| ContentReader
    ContentReader -->|Extracted Content| ResearchSynth

Diagram Explanation:

Summary

deep_search_r.json defines a multi-agent research orchestration framework that transforms complex user queries into high-quality strategic reports by leveraging specialized subagents for web discovery, extraction, and synthesis. It embodies a robust research methodology, adaptive execution, quality controls, and multilingual support, making it suitable for professional research tasks in business and consulting contexts.

End of Documentation