QUINTE

Five-Agent structured debate architecture for AI conclusion confidence.

Single-model AI hits a confidence ceiling. QUINTE breaks through — five independent agents debate your questions through structured rounds of analysis, cross-examination, and final verdict.

---

---

Architecture

Agent	Model	R1	R2	Strengths
Hermes	deepseek-v4-pro	✅	✅	Orchestration + final verdict
Claude Code	deepseek-v4-pro	✅	✅	Broadest coverage, structured reports
CodeWhale	deepseek-v4-pro	✅	✅	Deepest research, concurrency
OMP	deepseek-v4-pro	✅	✅	Full participant, LSP/DAP
Reasonix	deepseek-v4-pro	—	✅	Pure reasoning judge (R1 — limited)

R1: 4 agents. R2: 5 agents (+Reasonix). When Reasonix run mode supports tool calls, R1 expands to 5.

                             Hermes (Orchestrator + Participant)
                                              ▼
                  ┌─────────────────────────────────────────────────────────┐
                    Round 1         Round 1         Round 1        Round 1
                     Hermes         Claude         CodeWhale         OMP
                    (v4 xhigh)     (v4 max)        (v4 max)       (v4 xhigh)
                  └─────────────────────────────────────────────────────────┘
                                              ▼
                                  Hermes flags divergences
                                              ▼
              ┌───────────────────────────────────────────────────────────────────┐
                Round 2       Round 2       Round 2       Round 2       Round 2
                Hermes         Claude       CodeWhale     Reasonix        OMP
               (v4 xhigh)     (v4 max)      (v4 max)      (v4 max)     (v4 xhigh)
              └───────────────────────────────────────────────────────────────────┘
                                              ▼
                                     Hermes final verdict

Design Principles

• All DeepSeek v4-pro · Hermes/OMP xhigh, rest max · flash forbidden
• No degradation — all 5 agents must participate. Timeout → retry with smaller prompt, never skip.
• 3 rounds max — early consensus skips remaining rounds
• Cross-review is adversarial — review others, never yourself
• Terminal + background CLI — no delegate_task

Quick Start

git clone https://github.com/eric-stone-plus/quinte.git
cd quinte
open quinte.html          # Architecture visualization
bash quinte-demo.sh       # Simulate a debate round

Built With

QUINTE orchestrates five independent AI agents. None are developed by this project. Each is a standalone tool used as a debate participant.

Agent	Description	Repository
Hermes	Orchestrator + debater. Coordinates rounds and produces final verdict.	MIT
Claude Code	Anthropic's coding agent. Broadest coverage, structured reports.	MIT
CodeWhale	DeepSeek-native agent. Deepest research, concurrency analysis.	MIT
OMP	oh-my-pi fork. Full debate participant, LSP/DAP, 40+ providers.	MIT
Reasonix	DeepSeek-native reasoning. R2 pure judge, content-embedded.	MIT

All agents run on DeepSeek v4-pro. Hermes/OMP use xhigh, others use max. Flash is explicitly forbidden.

License

MIT — the protocol and orchestration layer. Individual agents carry their own licenses (see Built With).