Morlen

Morlen is a multi-agent decision engine (CFO, Ops, CX) built on Alibaba Cloud. Powered by Alibaba Cloud's DashScope API, Morlen uses Qwen agents to debate and manage your business operations natively.

Comment

Inspiration

In emerging markets, business doesn't happen on polished e-commerce dashboards; it happens in the chaotic, unstructured text of WhatsApp. I realized that giving a small business owner a standard AI chatbot wasn't enough. A generic LLM can answer FAQ questions, but it cannot evaluate profit margins, check real-time inventory, or analyze customer churn risk simultaneously. To run a business, you don't need a chatbot, you need an autonomous management team.

I built Morlen for Track 3: Agent Society. Morlen is a highly concurrent, multi-agent decision engine that routes complex customer interactions to a "Board of Directors" (CFO, Operations, and CX agents) who debate and resolve the issue natively using Alibaba Cloud's Qwen models.

What it does

When a complex customer request comes in (e.g., demanding a steep discount due to a late order), Morlen doesn't blindly reply. It triggers an asynchronous pipeline:

  • The CFO Agent (using qwen-max for deep reasoning) evaluates the strict financial margins of the requested discount.
  • The Ops Agent checks current stock availability and logistics constraints.
  • The CX Agent analyzes the customer's lifetime value and churn risk.

The three agents actively debate the outcome in the background. Once they reach a consensus, the agreed-upon action is executed, or the customer receives a unified, highly-contextual reply.

The Killer Feature: Solving "Agentic Abstention"

Top researchers recently warned about "Agentic Abstention": the phenomenon where multi-agent systems hallucinate or get trapped in infinite API loops when they disagree.

Morlen solves this natively at the architectural level. I built a strict, 1-turn Conflict Resolution Synthesizer. If the CFO says 'No' and Ops says 'No Stock', the Synthesizer deterministically overrides the loop, kills the automated request to prevent wasted compute, and escalates the thread to the human business owner via the Signal Inbox. The owner sees the exact reasoning trace of the AI debate before stepping in.

How I built it

Morlen is built on a highly reliable, deterministic orchestration stack:

  • Intelligence Layer: Built natively on Alibaba Cloud DashScope. I dynamically route low-latency tasks to qwen-plus and reserve qwen-max for the deep, multi-agent reasoning board.
  • Orchestration Layer: A Python backend utilizing asyncio.gather for simultaneous agent invocation.
  • Deployment: Architected for Alibaba Cloud ECS (Elastic Compute Service) using Docker containerization, with native Bash deployment scripts (deploy_alibaba_ecs.sh).
  • Frontend: A Next.js and React dashboard that streams the multi-agent consensus live to the business owner.

Challenges I ran into

Handling the hardware and latency constraints of invoking three distinct LLM agents simultaneously was a significant challenge. Initially, running the agents sequentially caused unacceptable response times for a live chat interface. I solved this by strictly modularizing the prompts and aggressively executing them concurrently via our Python backend, utilizing DashScope's high-throughput endpoints to return the final Synthesized consensus in milliseconds.

Accomplishments that i'm proud of

  1. Flawless Multi-Agent Orchestration: Successfully isolating the CFO, Ops, and CX roles without letting their distinct system prompts bleed into one another.
  2. Deterministic Safety: Engineering the Conflict Resolution Synthesizer to guarantee that Morlen will never get stuck in an infinite API loop, saving compute costs.
  3. Alibaba Cloud Native: Building the entire backend orchestration directly on DashScope APIs rather than relying on heavy, third-party wrapper libraries.

What I learned

I learned that the future of reliable AI engineering isn't just about giving a single model a larger context window—it's about orchestration. By breaking a problem down and assigning it to highly specialized, distinct agents, and forcing them to debate, the underlying Qwen models produced vastly superior, enterprise-grade reasoning.

What's next for Morlen

I want to take Morlen from a localized management team to a true global commerce engine. My next milestones include integrating real-time accounting APIs into the CFO agent, and expanding the deployment cluster on Alibaba Cloud to support thousands of simultaneous multi-agent debates for small businesses worldwide.

Built With

Share this project:

Updates