Deep Research

# Deep Research


Conduct thorough, multi-source deep research using subagents for parallel extraction and a coordinator for synthesis. This is the best approach: one agent plans and synthesizes, subagents extract from individual sources in parallel.


## Table of Contents

- [When to Use](#when-to-use)
- [Triggers](#triggers)
- [Architecture](#architecture)
- [Methodology](#methodology)
- [Key Findings](#key-findings)
- [Conflicting Viewpoints](#conflicting-viewpoints)
- [Confidence Assessments](#confidence-assessments)
- [Gaps & Open Questions](#gaps-and-open-questions)
- [Nonsense Check](#nonsense-check)
- [Pitfalls](#pitfalls)
- [Cross-Skill Hints](#cross-skill-hints)

## When to Use

Use when asked to research a topic, compare options, investigate a question, analyze a problem, or find supporting evidence. Also when you need reliable information to inform a decision, write a document, or validate a claim.

## Triggers

"research", "deep dive", "investigate", "analyze", "look into", "find information about", "I need to understand", "compare options", "what does the evidence say", "do some research", "find studies on", "gather information", "comprehensive analysis", "literature review", "systematic review", "evidence synthesis"

## Architecture

```
Coordinator (you)
├── Phase 1: Formulate queries            (single thread)
├── Phase 2: Source selection             (single thread)
├── Phase 3: Parallel extraction          (spawn subagents)
│   ├── Subagent 1: Source A extraction
│   ├── Subagent 2: Source B extraction
│   ├── Subagent 3: Source C extraction
│   └── ...
├── Phase 4: First synthesis              (merge subagent findings)
├── Phase 5: Gap analysis & second pass   (new queries + sources +subagents if needed)
└── Phase 6: Final synthesis & validation (single thread)
```

This parallel pattern is 3-5x faster than sequential extraction and produces better synthesis because each subagent focuses on one source deeply. Google DeepMind's 2025 study of 180 agent system configurations confirmed that parallel extraction with a coordinator (orchestrator topology) outperforms sequential and linear chain topologies for information retrieval tasks (arXiv:2512.08296). Anthropic's engineering team also validates the orchestrator-worker pattern as the preferred architecture for multi-agent research systems (Anthropic, 2025).

## Methodology

### Phase 1: Query Formulation