Deep Researching

Overview

Multi-agent research: decompose a question into sub-topics, dispatch parallel researchers, synthesize findings into a cited report, and archive all sources locally.

When to Use

• User asks a broad or complex research question
• Task requires synthesizing information from many sources
• User wants citations and verifiable sources
• User explicitly asks for "deep research" or a "research report"

Don't use for: Quick factual lookups, code questions, tasks that don't need web research.

Process

digraph research {
    rankdir=TB;
    decompose [label="1. Decompose query\ninto 3-5 sub-topics" shape=box];
    mkdir [label="Create output dir\nresearch-{slug}-{timestamp}/" shape=box];
    broad [label="2. Dispatch 3-5\nresearch-worker agents\n(PARALLEL)" shape=box style=bold];
    synthesize [label="3. Synthesize findings\nIdentify gaps & conflicts" shape=box];
    gaps [label="Gaps found?" shape=diamond];
    followup [label="4. Dispatch 1-3 more\nresearch-workers\n(PARALLEL)" shape=box style=bold];
    archive [label="5. Dispatch source-archiver\n(BACKGROUND)" shape=box style=bold];
    report [label="6. Write report.md\nusing report-format template" shape=box];
    done [label="Present report path\nto user" shape=box];

    decompose -> mkdir -> broad -> synthesize -> gaps;
    gaps -> followup [label="yes"];
    gaps -> archive [label="no"];
    followup -> archive;
    archive -> report -> done;
}

Execution Details

Phase 1: Query Decomposition

Analyze the research question. Break it into 3-5 independent sub-topics that together cover the full scope. Create output directory:

mkdir -p ./research-{topic-slug}-{YYYYMMDD-HHMMSS}/sources

Phase 2: Broad Research

Dispatch 3-5 deep-research:research-worker agents IN PARALLEL via Task tool. Each gets:

• The overall research question (for context)
• ONE specific sub-topic to investigate
• Instruction to return structured findings with URLs

Phase 3: Synthesis & Gap Analysis

Read all worker results. Across all findings:

• Merge and deduplicate sources
• Assign sequential source numbers [1], [2], ... to all unique URLs
• Identify gaps: sub-topics with thin coverage
• Identify conflicts: sources that disagree
• Identify promising leads not yet followed

Phase 4: Targeted Follow-up

If gaps exist, dispatch 1-3 more deep-research:research-worker agents with specific, narrow queries informed by Phase 2 findings. Skip if coverage is sufficient.

Phase 5: Source Archival

Collect all unique source URLs with their assigned numbers. Dispatch deep-research:source-archiver agent IN BACKGROUND (run_in_background: true) with the full URL list and output directory path. Do not wait for it to finish before writing the report.

Phase 6: Report Generation

Write report.md to the output directory. Follow the template in references/report-format.md:

• Executive summary answering the user's question
• Thematic sections with inline citations [1]
• Key findings list
• Limitations & open questions
• Sources table with links to local archived copies

Present the output directory path to the user when complete.