Community

Blog Events Webinars Tutorials Forum
  1. Blog
  2. Events
  3. Webinars
  4. Tutorials
  5. Forum
Create Account
×
Community Blog Tarmac: Five AI Agents Negotiate the Storm, and One Mediator Signs the Ruling

Tarmac: Five AI Agents Negotiate the Storm, and One Mediator Signs the Ruling

This article introduces Tarmac, an auditable multi-agent AI system that negotiates fair passenger rebookings during airline disruptions.

cover

Track Winner — Track 3: Agent Society, Global AI Hackathon Series with Qwen Cloud

Project Introduction

A storm closes an airport. One hundred and eighty passengers are displaced; only 139 legal seats remain. Somewhere in the crowd are a medical courier carrying time-critical supplies, an unaccompanied minor, and a passenger who needs a wheelchair. A single optimization algorithm, asked to fill seats efficiently, will quietly strand exactly the people who can least afford to be stranded — and then be unable to explain why.

Tarmac is the alternative. It's a multi-agent society for airline irregular operations: when a storm cancels flights, five Qwen agents with genuinely opposing goals negotiate passenger rebookings on a shared seat ledger, and a Qwen mediator resolves deadlocks with signed rulings that cite the actual regulation. Trade-offs are argued out in the open, not hidden inside a black-box objective function.

Project Value

The value is measured, and the numbers are striking. On a frozen storm scenario — 180 displaced passengers, 139 legal seats, run across 10 random seeds — a single planner stranded 17 protected passengers. The agent society stranded just 3. It met 100% of special-needs SLAs, versus 0% for the single planner, and it committed zero crew-duty violations. Every decision can be audited and replayed after the fact.

The deeper point is about _legibility_. A lone optimizer can't tell you why it left the wheelchair user on the tarmac. Tarmac can, because the decision was a negotiation between named interests — rebooking, crew legality, gates and ground, hotels, and passenger advocacy — each arguing its case on the record. When something has to give, you can see exactly what gave and why.

Solution

Five Qwen agents — Rebooking, Crew-Legality, Gate/Ground, Hotel, and Passenger-Advocate — place revocable claims on a row-locked seat ledger. Contested seats go to sealed-bid rounds, so no agent can bully another by watching its moves. When the agents deadlock, a Duty-Manager mediator steps in and issues an Ed25519-signed ruling that cites the relevant rule — a cryptographic signature on the tie-break.

Every step is written to a hash-chained log that replays byte for byte, so any outcome can be independently verified after the fact. A real end-to-end live run is committed to the repository (not a mock — an actual run on live models, 324 log entries), and the project carries 330 tests at 100% coverage. It's an agent society built with the rigor of an auditing firm.

Highlights

The five negotiating agents run on qwen3.7-plus through Qwen Cloud's OpenAI-compatible endpoint (Model Studio / DashScope International), while the Duty-Manager mediator runs on qwen3.7-max for the rare deadlock rulings. That split is deliberate economics: qwen3.7-plus holds persona and tool discipline across roughly 60 turns at low cost, which makes the society affordable to benchmark over many seeds, while the few high-stakes rulings get the strongest model available.

The orchestrator is deployed on Alibaba Cloud Function Compute (python3.10, ap-southeast-1). Its /verify endpoint re-runs the full invariant suite against a live Qwen run byte for byte in the cloud, /run executes a fresh society on demand, and /health serves as the liveness probe. The result is an agent society you can't just trust — you can check.

Alibaba Cloud Products Used

About the Team

Edy Cu is a solo developer based in Jakarta, Indonesia, who builds at hackathons almost every week and maintains 160+ public repositories — most of them AI agents that run on real model calls, are deployed on public endpoints, and come with reproducible benchmarks. Tarmac won Track 3: Agent Society and is MIT-licensed. Its negotiation core is domain-agnostic; the repo even includes a short non-airline example (meeting-room allocation), and Edy is developing it into a standalone open-source protocol for auditable agent societies on Qwen.

Explore the project: https://github.com/edycutjong/tarmac

Demo: https://youtu.be/cOfFwumzG7k

This project was developed as part of the Global AI Hackathon Series with Qwen Cloud.

AI Qwen AI Hackathon Qwen Models Multi-Agent Systems AI Negotiation Auditable AI
0 0 0
Share on

Read previous post:

Rexgent: One Idea In, a Whole Mini-Drama Out

Read next post:

Sentinel Qwen Ensemble: The Autonomous Forensic Investigator That Can't Bluff

Alibaba Cloud Project Hub

140 posts | 29 followers

You may also like

Comments

Alibaba Cloud Project Hub

140 posts | 29 followers

Related Products

  • Token Plan

    Build more, spend less. One plan, every modality.

    Learn More

  • QwenWork

    QwenWork is dedicated to helping employees strengthen their professional competitiveness in the AI era and to enabling enterprises to improve organizational effectiveness.

    Learn More

  • Qwen

    Full-range, open-source, multimodal, and multi-functional

    Learn More

  • Alibaba Cloud Model Studio

    A one-stop generative AI platform to build intelligent applications that understand your business, based on Qwen model series such as Qwen-Max and other popular models

    Learn More

More Posts by Alibaba Cloud Project Hub

See All