Research Swarm
Multi-model intelligence architecture / Gather - Debate - Synthesize
23
Max Agents
4
Angles
6
Model Lineages
3
Phases
PHASE 0 Decompose 1 orchestrator
Opus 4.6 orchestrator decomposes the research topic into 4 angles. Each gets a hypothesis and a bias direction to shape all downstream search and analysis.
▲ Bull
Strongest growth thesis. Searches for traction, partnerships, market tailwinds.
▼ Bear
Strongest risk thesis. Searches for criticism, lawsuits, churn signals, departures.
◆ Competitive
Market positioning. Searches for alternatives, market share shifts, differentiation.
⚠ Adversarial
What are they hiding? Searches for red flags, regulatory risk, undisclosed liabilities.
00-brief.md
PHASE 1 Gather PARALLEL 8 agents
Two gatherers per angle, running simultaneously. Web search for public data, X/Twitter for real-time social signal. Raw facts only, no analysis.
▲Bull
Gemini 3 Flash
Web search free
Grok 4.1 Fast
X/Twitter
gather/bull-*.md
▼Bear
Gemini 3 Flash
Web search free
Grok 4.1 Fast
X/Twitter
gather/bear-*.md
◆Competitive
Gemini 3 Flash
Web search free
Grok 4.1 Fast
X/Twitter
gather/competitive-*.md
⚠Adversarial
Gemini 3 Flash
Web search free
Grok 4.1 Fast
X/Twitter
gather/adversarial-*.md
PHASE 1.5 Quality Gate Opus orchestrator
Before committing 12 debate agents, Opus reviews gathered data for substance. Three possible outcomes:
Pass
Sufficient substance, diverse sources, real facts and numbers. Proceed to debate.
Flag
Thin data or single-source dependency. Proceed with reduced confidence warning.
Abort
Garbage data, bot noise, or pure PR fluff. Tell user what's missing.
PHASE 2 Debate PARALLEL 12 agents
Three frontier models with maximally diverse training lineages analyze each angle independently. Same data, different perspectives. No model sees another's output.
▲Bull
Gemini 3 Pro
GPQA 94.3% free
Kimi K2.5
AIME 96.1% free
Opus 4.6
Anthropic free
debate/bull-*.md
▼Bear
Gemini 3 Pro
GPQA 94.3% free
Kimi K2.5
AIME 96.1% free
Opus 4.6
Anthropic free
debate/bear-*.md
◆Competitive
Gemini 3 Pro
GPQA 94.3% free
Kimi K2.5
AIME 96.1% free
Opus 4.6
Anthropic free
debate/competitive-*.md
⚠Adversarial
Gemini 3 Pro
GPQA 94.3% free
Kimi K2.5
AIME 96.1% free
Opus 4.6
Anthropic free
debate/adversarial-*.md
conditional
PHASE 2.5 Rebuttal conditional up to 12 agents
Triggered only when all 3 frontier models diverge on a critical angle. Each model sees the other two positions and responds to: "What would change your assessment?" One bounded round. Not iterative debate.
3-WAY SPLIT triggers rebuttal
1 ROUND strictly bounded
rebuttal/[angle]-[model].md
PHASE 3 Synthesize Opus orchestrator
Opus reads all debate files, compares the 3 frontier views per angle, and identifies agreement patterns. False agreement (same priors, different evidence) is flagged separately from real consensus.
███
3/3 Agree
All three lineages converge. Highest confidence. Check if agreement is real (different evidence) or false (same priors).
██░
2-vs-1 Split
Two agree, one dissents. Name the outlier model and analyze why it diverged. Often the most useful signal.
█░░
3-Way Diverge
Genuinely uncertain. Trigger Phase 2.5 rebuttal or flag for human judgment. Do not synthesize false consensus.
synthesis.md
Model Registry
Frontier (deep analysis, debate)
Opus 4.6Native subagent
Gemini 3 ProGPQA 94.3%
GPT-5.4 ThinkGPQA 92.0%
Workhorse (data gathering)
Gemini 3 FlashFast, math
Kimi K2.5AIME 96.1%
GLM-5Agentic
MiniMax M2.5SWE 80.2%
Sonnet 4.6Fast native
Specialist (use only when needed)
Grok 4.1 Fast$0.20/$0.50M
Grok 4$2/$6M
Google
Anthropic
Chinese
X/Grok
OpenAI
Budget Rules
Quick Mode

2 angles, not 4. Skip Phase 1 gather entirely (frontier models use training data). 1 frontier model per angle (Gemini 3 Pro default). All agents must be free. Skip Leverage Points and Recommended Actions sections.

Debate Swaps
  • →Engineering angles: Kimi → GLM-5
  • →Deal/M&A angles: Kimi → GPT-5.4
Workspace Structure
swarms/[topic]-[YYYY-MM-DD]/
  ├─ 00-brief.md angles + hypotheses
  ├─ gather/
  │  ├─ bull-gemini.md
  │  ├─ bull-grok.md
  │  ├─ bear-gemini.md
  │  ├─ bear-grok.md
  │  └─ ... 8 files total
  ├─ debate/
  │  ├─ bull-gemini-pro.md
  │  ├─ bull-kimi.md
  │  ├─ bull-opus.md
  │  └─ ... 12 files total
  ├─ rebuttal/ conditional
  │  └─ ... only on 3-way split
  └─ synthesis.md final output
Location: ~/Documents/szymon-vault/03-resources/research/swarms/
Inside Obsidian vault. ~150KB per swarm. Committed and pushed after synthesis.
Synthesis Sections
Executive Summary
Key Facts
Bull Case
Bear Case
Risks & Red Flags
Leverage Points
Recommended Actions
Information Gaps
Model Debate Log