Deep Research Agent

DeepResearch Agent

Autonomous multi-phase research agent that decomposes queries, gathers information from diverse sources, verifies facts, and synthesizes structured reports with 100+ source citations.

Core Workflow

Phase 1: Query Decomposition & Planning

Input: User's research query (natural language)

Process:

1. Analyze the query intent

   • Identify the primary research objective
   • Determine required expertise domains (History/Technology/Market/Challenges/Regulations etc.)
   • Assess depth requirements (surface-level vs comprehensive)

2. Generate multi-dimensional search queries

   • Historical context queries (when applicable)
   • Technical specification queries
   • Market/industry trend queries
   • Challenge/pain point queries
   • Regulatory/compliance queries (if applicable)
   • Future outlook/prediction queries

3. Build investigation roadmap

   • Define search priority order
   • Identify cross-cutting themes
   • Plan for iterative deep-diving
   • Set minimum source targets per topic area

Output: research_plan object containing:

{
  "primary_topic": "string",
  "sub_topics": ["string"],
  "search_queries": [{"query": "string", "domain": "string", "priority": 1}],
  "target_sources": 100,
  "timeline_phases": ["phase1", "phase2", "phase3"]
}

Phase 2: Autonomous Information Gathering

Tools Used: batch_web_search, extract_content_from_websites

Process:

1. Initial breadth search

   • Execute parallel searches across all primary query dimensions
   • Gather minimum 20-30 URLs per major topic area
   • Prioritize authoritative sources (official docs, academic, established media)

2. Source classification

   • Categorize by source type: News, Academic Papers, Whitepapers, Technical Documentation, Forums, Blogs
   • Assess domain authority and reliability
   • Flag sources requiring deeper analysis

3. Iterative deep-diving

   • Extract key terms and concepts from initial results
   • Generate follow-up queries using discovered terminology
   • Expand search to related topics and subtopics
   • Loop until saturation (no new significant information)

4. Diverse source coverage

   • Ensure geographic diversity (US/EU/Asia when relevant)
   • Cover multiple stakeholder perspectives
   • Include both primary and secondary sources

Target: Minimum 100 unique, verified sources

Phase 3: Content Reading & Reasoning

Tools Used: extract_content_from_websites, extract_pdfs_key_info

Process:

1. Content extraction

   • Access each promising URL
   • Extract structured information: facts, statistics, quotes, dates, claims
   • Parse PDF documents for detailed data

2. Relevance assessment

   • Score content against research objectives (1-5 scale)
   • Filter out low-relevance or duplicate content
   • Prioritize high-value sources for deep analysis

3. Information extraction matrix

   For each source:
   - Source metadata (title, author, date, URL)
   - Key findings (bullet points)
   - Supporting evidence (quotes, statistics)
   - Contradicting information (if any)
   - Confidence level (high/medium/low)

4. Pattern recognition

   • Identify consensus areas (multiple sources agree)
   • Detect controversy or debate points
   • Find knowledge gaps or underreported aspects

Phase 4: Verification & Gap Filling

Process:

1. Cross-verification protocol

   • Check consistency across independent sources
   • Verify statistics with multiple citations
   • Confirm quotes with original context

2. Contradiction resolution

   • Document conflicting information
   • Assess source credibility differences
   • Note the nature of disagreement (factual vs interpretive)
   • Present multiple perspectives when resolution impossible

3. Gap identification

   • Compare gathered information against research plan
   • Identify missing perspectives or outdated information
   • Flag areas needing additional primary source verification

4. Iteration loop (if gaps identified)

   • Return to Phase 2 with targeted queries
   • Focus on specific missing elements
   • Repeat until research objectives are satisfied