Deep Research

Turn any question into multi-agent exploration.

                              ┌─────────────────┐
                              │  YOUR QUESTION  │
                              └────────┬────────┘
                                       │
                                       ▼
                              ┌─────────────────┐
                              │  ORCHESTRATOR   │
                              │    (Claude)     │
                              └────────┬────────┘
                                       │
                    ┌──────────────────┼──────────────────┐
                    │                  │                  │
                    ▼                  ▼                  ▼
             ┌────────────┐     ┌────────────┐     ┌────────────┐
             │ Researcher │     │ Researcher │     │ Researcher │
             │    (1)     │     │    (2)     │     │   (...)    │
             └─────┬──────┘     └────────────┘     └─────┬──────┘
                   │                                     │
              ┌────┴────┐                           ┌────┴────┐
              ▼         ▼                           ▼         ▼
           ┌─────┐   ┌─────┐                     ┌─────┐   ┌─────┐
           │ Sub │   │ Sub │        ...          │ Sub │   │ Sub │
           └─────┘   └─────┘                     └─────┘   └─────┘

                         Agents spawn as many
                       sub-agents as needed (∞)

                                   │
                                   ▼
                              ┌─────────┐
                              │SYNTHESIS│
                              └─────────┘

Multi-agent exploration

./deep-research.sh "Why do smart people make bad decisions?"

Each agent autonomously decides: break it down further or answer directly. Researchers run in parallel. Unlimited depth. Results flow back up into synthesis.

---

Quick Start

git clone https://github.com/Cranot/deep-research.git
cd deep-research
chmod +x deep-research.sh

# Run it
./deep-research.sh "Your question here"

Requires: Claude Code CLI

---

Example Reports

Real outputs from actual runs:

Question	Config	Result
What makes great software architecture?	sonnet + haiku	6 dimensions identified
What is the key to learning fast?	sonnet + haiku	5-principle framework
Why do people self-sabotage?	sonnet	Psychological deep-dive
What makes people buy vs just admire?	opus	Purchase psychology

---

Options

# Default: Opus orchestrator + Sonnet researchers
./deep-research.sh "Your question"

# Fast & cheap: Sonnet + Haiku
./deep-research.sh -m sonnet -r haiku "Your question"

# With web search (for current events)
./deep-research.sh --web "Latest developments in AI agents"

# Maximum depth: Sonnet + Sonnet (researchers can spawn more)
./deep-research.sh -m sonnet -r sonnet "Complex topic"

Flag	What it does	Default
-m, --model	Orchestrator model	opus
-r, --researcher	Researcher model	sonnet
-w, --web	Enable web search	off

Note: Haiku researchers answer directly and won't spawn sub-researchers. Use -r sonnet if you want deeper recursion.

---

Model Combinations

Config	Depth	Speed	Cost	Best for
-m opus -r sonnet	Deep	Slow	$$$	Best quality
-m sonnet -r sonnet	Deep	Medium	$$	Good balance
-m sonnet -r haiku	Shallow	Fast	$	Quick answers

Costs vary by question complexity and depth. Budget accordingly.

---

Output

Reports save automatically to:

reports/YYYY-MM-DD-your-question-slug/SYNTHESIS.md

---

How It Works

The magic is one flag: --allowedTools "Bash(claude:*)"

This lets Claude spawn more Claude instances. Each instance gets the same "DNA":

Before answering, ask: are there multiple angles worth exploring?
- If YES → spawn sub-agents IN PARALLEL, wait for all, synthesize
- If NO → answer directly
Bias toward exploring. Go deep, not shallow.

Each agent spawns researchers for different angles:

claude -p "sub-question-1" --model sonnet ...
claude -p "sub-question-2" --model sonnet ...
claude -p "sub-question-3" --model sonnet ...
# collect all results, then synthesize

Recursive by nature. Each agent can spawn more. The tree grows until questions become atomic.

---

Built by

AgentsKB - Pre-researched knowledge for AI agents

License

MIT