三人行 · sanrenxing

三人行，必有我师。 — Among any three, one can be my teacher.

Divergent, open-mind AI discussion. Three AI seats with complementary lenses debate a question — not to vote a winner, but to open the possibility space. Each round, every seat reacts to the others (spark / tension / gap) and generates new angles. A curator then lays out a possibility map of distinct branches — each with its premise, cost, and concrete next step. You choose the path.

Provider-agnostic: every seat is just an OpenAI-compatible endpoint, so mix any models you like (OpenAI, DeepSeek, Moonshot/Kimi, a local vLLM/Ollama, …).

▶ Try it live: https://taolab.tail0ea5ac.ts.net/orchestra/ (hosted by Tao Lab)

_{Illustrative example — three seats open distinct angles, the curator lays out branches A/B/C with premise, cost, and a concrete next step.}

Ships in three forms:

• MCP server — trio_round + fanout tools for any MCP host (Claude Code, etc.)
• Web demo — a zero-framework single page with live streaming (web/)
• CLI / library — python -m sanrenxing "…" or import sanrenxing

---

Why divergent, not convergent

Most "multi-agent debate" tools push agents to argue until they converge on one answer. 三人行 does the opposite on purpose:

	convergent debate	三人行 (divergent)
goal	pick the winner	open more branches
seats	same lens, vote	complementary lenses
cross-talk	rebut / eliminate	yes-and / generate
output	one recommendation	a map, human picks

Convergence is treated as a smell. The three default seats are seeded to disagree in kind — evidence (证据席), critique (批判席), and the unconventional angle (发散席) — so the discussion stays heterogeneous. Convergent mode (vote/rank) is opt-in: just ask the host to decide.

Architecture

question
   │
   ├─► Seat 1 ─┐
   ├─► Seat 2 ─┼─ round 1 (parallel, distinct angles)
   ├─► Seat 3 ─┘
   │      ▼  each seat reads the others' prior round, reacts + branches
   ├─► round 2 … N  (mutual-evaluation = the interaction engine)
   │
   └─► Curator ─► possibility map { direct, overview, branches[premise/cost/next], more }

• sanrenxing/config.py — load seats from env (SEATn_*).
• sanrenxing/llm.py — minimal OpenAI-compatible client (block + stream).
• sanrenxing/discussion.py — seat prompts, rounds, curator. The methodology.
• mcp_server/server.py — trio_round, fanout MCP tools.
• web/server.py + web/static/index.html — stdlib HTTP + SSE demo.

Quick start

git clone https://github.com/taoyongac/sanrenxing
cd sanrenxing
pip install -r requirements.txt
cp .env.example .env        # then edit: set SEAT1/2/3 model + base_url + api_key

CLI

set -a; source .env; set +a
python -m sanrenxing "短端粒在衰老中是刹车还是油门？"
python -m sanrenxing -r 3 "your question"     # 3 rounds

Web demo

set -a; source .env; set +a
python web/server.py        # → http://127.0.0.1:8030

MCP server — add to your host config (e.g. Claude Code ~/.claude.json), making sure the SEATn_* vars are in its environment:

{
  "mcpServers": {
    "sanrenxing": {
      "command": "python",
      "args": ["/abs/path/to/sanrenxing/mcp_server/server.py"],
      "env": { "SEAT1_MODEL": "gpt-4o", "SEAT1_API_KEY": "sk-...",
               "SEAT2_MODEL": "deepseek-chat", "SEAT2_BASE_URL": "https://api.deepseek.com", "SEAT2_API_KEY": "sk-...",
               "SEAT3_MODEL": "moonshot-v1-32k", "SEAT3_BASE_URL": "https://api.moonshot.cn/v1", "SEAT3_API_KEY": "sk-..." }
    }
  }
}

Then the host can call trio_round(seat1_prompt, seat2_prompt, seat3_prompt, round_num) once per round (embedding the prior transcript into each prompt for rounds 2+), and fanout(tasks) for parallel independent angles. The host stays the arbiter — the tools never merge or pick a winner.

Configuration

All via environment (see .env.example). Per seat n ∈ {1,2,3}:

var	meaning	default
SEATn_MODEL	model id (required to enable the seat)	—
SEATn_BASE_URL	OpenAI-compatible base url	OPENAI_BASE_URL or OpenAI
SEATn_API_KEY	api key	OPENAI_API_KEY
SEATn_NAME	display label	证据席 / 批判席 / 发散席
SEATn_PERSONA	one-line lens steering the angle	sensible per-seat default

CURATOR_MODEL/BASE_URL/API_KEY override the curator (defaults to Seat 1). The web demo also reads SANRENXING_HOST/PORT/ROUNDS and optional SANRENXING_USER/PASS (HTTP Basic Auth for a shared deployment).

Notes

• Seats are pure reasoning (no tools/web). Embed any needed facts in the question; tool/search augmentation is intentionally left as an extension.
• A single global lock serializes web discussions (concurrency = 1) so a shared demo never piles up parallel runs.

---

From the Tao Lab

Built and used at Tao Lab, School of Life Sciences, Yunnan University (云南大学 · 陶勇课题组) — epigenetics, aging, cancer, and AI-for-Science. 三人行 is one of the lab's open tools for turning a hard scientific question into a map of testable directions.

🔗 Lab site: https://taolab.tail0ea5ac.ts.net/ ▶ Live 三人行: https://taolab.tail0ea5ac.ts.net/orchestra/

License

MIT © 2026 Yong Tao (Tao Lab, Yunnan University). See LICENSE.