Building a Multi-Agent Society with Qwen Cloud — My Hackathon Journey

#alibabachallenge #ai #agents #hackathon

When I first read the Track 3 description — "Design a multi-agent collaboration system where multiple Agents with distinct capabilities work together through task division, dialogue, and negotiation" — I knew I had to build something beyond the typical "ask an AI and get an answer" pattern.

What if agents could debate with each other? What if they could disagree, negotiate, and reach consensus — just like a real team of experts?

That's how Qwen Council was born.

The Problem

Single-agent AI code review has a fundamental limitation: one model, one perspective. It might catch SQL injection but miss architectural debt. It might flag a performance issue but overlook accessibility problems.

In the real world, code reviews are done by teams — security engineers, architects, QA specialists. Each brings a different lens. Why should AI be any different?

The Architecture

Qwen Council is a multi-agent system with 14 specialised agents running on Qwen Cloud, deployed on Alibaba Cloud ECS.

Two Modes

Code Review — 6 agents with distinct expertise (Security, Architecture, Quality, Performance, UX, Vision) review code through 4 structured debate rounds: individual analysis, cross-debate, refinement, and negotiation.

General Chat — 8 personality-based agents (inspired by Feynman, Torvalds, Socrates, Harari, Miyazaki, Jung, Sun Tzu, and Franklin) answer any question. Each agent has a strict domain boundary — they decline out-of-scope questions. A router classifies the question and activates only the 1-3 most relevant agents.

The Communication Protocol

The key innovation is an inter-agent protocol based on cognitive linguistics:

Inverted Pyramid: Each finding starts with the conclusion first — one line any agent can grasp in 1 second.
Given-New: In debate rounds, agents explicitly reference each other's findings before adding new evidence.

3-Level Memory

Working memory (volatile) holds the current session. Episodic memory (PostgreSQL) stores past sessions with a forgetting curve that decays unused data. Semantic memory (PostgreSQL + pgvector) consolidates recurring patterns as vector embeddings.

The Results

I ran a benchmark comparing a single generalist agent vs the full council on the same code:

Metric	Single-Agent	Multi-Agent	Change
Total findings	12	127	+958%
Categories covered	6/6	6/6	100% overlap
Unique findings	0	115	+115

The multi-agent system found 10.6x more issues while preserving everything the single agent caught.

Deploying on Alibaba Cloud

The entire stack runs on a single Alibaba Cloud ECS instance (2 vCPU, 4 GB RAM) in the Singapore region. Three Docker containers handle everything:

Nginx serves the React frontend and routes API calls to the backend
FastAPI orchestrates the agents and manages the memory system
PostgreSQL with pgvector stores both session data and semantic embeddings

The backend connects to Qwen Cloud via its OpenAI-compatible API, using three models: qwen3-coder-plus for code analysis, qwen-vl-plus for visual inspection, and text-embedding-v3 for semantic memory embeddings.

Why Alibaba Cloud? The ECS + Docker Compose setup gave me a simple, production-ready deployment with zero orchestration overhead. Qwen Cloud's native integration meant low-latency API calls from the ECS instance. And pgvector let me run hybrid relational + vector search in a single database — no separate vector DB needed.

What I Learned

Agents need boundaries. Initially, every agent answered every question with equal competence. Adding explicit domain restrictions made the system much more realistic.
Routing matters. A keyword-based classifier routes questions to 1-3 relevant agents. Getting this right is crucial — misrouting wastes tokens and dilutes quality.
Memory is hard. The forgetting curve required careful tuning. Too aggressive and sessions disappear; too lenient and the database fills up.
SSE streaming transforms UX. Watching agents respond in real-time is dramatically more engaging than waiting for a single response.