Multi-Agent Brainstorm
Overview
Structured brainstorming that combines clarifying dialogue with parallel expert sub-agents. The goal: understand the problem first, then get diverse expert perspectives, then synthesize.
Process
Phase 1: CLARIFY → Understand the topic (2-3 targeted questions) Phase 2: MAP → Select relevant expert agents Phase 3: EXPLORE → Spawn agents in parallel Phase 4: SYNTHESIZE → Combine and present findings
Phase 1: Clarify the Topic
Before spawning any agent, understand what the user actually needs.
Ask 2-3 targeted questions (one at a time, prefer multiple choice):
- What is the core problem or goal?
- What constraints exist? (tech stack, budget, timeline, existing code)
- What does success look like?
Rules:
- One question per message
- Prefer multiple choice via AskUserQuestion when possible
- Max 3 questions — don't interrogate, get enough context to spawn useful agents
- If the topic is already clear and specific, skip to Phase 2
Phase 2: Map to Expert Agents
Select 2-4 relevant agents from this mapping:
| Domain | Agent Type |
|---|---|
| AI/ML Research | research-synthesizer |
| Creative / Image Gen | midjourney-expert |
| Finance/Crypto/Tax | finance-advisor |
| Prompt Engineering | prompt-engineer |
| Architecture/Design | feature-dev:code-architect |
| Code Review | feature-dev:code-reviewer |
| General research | general-purpose |
Selection criteria:
- Pick agents that bring different perspectives on the topic
- At least one agent should challenge the obvious approach
- For cross-domain topics, pick agents from different domains
Phase 3: Explore via Parallel Agents
Spawn all selected agents simultaneously using the Agent tool.
Each agent prompt MUST include:
Brainstorm: [topic with context from Phase 1] Context: [key constraints and goals identified] Your task: 1. Propose 2-3 approaches from your domain expertise 2. For each approach: trade-offs, risks, and when it works best 3. Justify every recommendation with sources (docs, papers, experience) 4. State confidence level (HIGH/MEDIUM/LOW) for each claim 5. Flag anything you cannot verify 6. Highlight what you'd do differently from the obvious approach
Important:
- Use
model: "opus"for complex/architectural topics - Use
model: "sonnet"for straightforward domain questions - All agents run in parallel — never sequential
Phase 4: Synthesize
After all agents return, present a unified synthesis:
markdown## Synthesis: [Topic] ### Consensus [What multiple agents agree on — strongest signal] ### Divergent Views [Where agents disagree, with rationale from each side] ### Recommended Approach [Best path forward, justified by agent findings] [State which agent perspectives informed this recommendation] ### Trade-offs to Consider [Key decisions the user still needs to make] ### Sources [All sources cited by agents, deduplicated] ### Confidence: [LEVEL] [Based on source agreement and verification quality]
Edge Cases
Topic too vague
Ask one clarifying question. If still vague after 2 questions, brainstorm at the level of abstraction given.
No obvious agent match
Use general-purpose agents with specific domain prompts. Always spawn at least 2.
User wants to go deeper
After synthesis, offer to spawn additional agents on specific sub-topics that emerged.
Quick brainstorm
If user says "quick" or the topic is narrow, skip Phase 1 and spawn 2 agents directly.
