Award-winning project in the Global AI Hackathon Series (organized by Qwen Cloud)

Project Overview

A disaster occurs in a virtual city. Various missions unfold on the map: a flooded residential area, a collapsed school building with disaster victims trapped inside, a hospital whose generator fuel has run out, and so on. To respond to this situation, an “agent society” composed of Qwen agents with six clearly defined roles (incident commander, medical response, firefighting, rescue, infrastructure, and public relations) is established. This is “Aftershock.” The project is a deterministic disaster-response simulator and, at the same time, a real-time observation platform for the Qwen agent society. Its raison d'être is to answer the fundamental question that the industry as a whole has vaguely avoided—“When multiple agents are deployed together, does their coordination actually work?”

The Project's Value

Coordination among multiple agents is easy to claim, but extremely difficult to demonstrate rigorously. Many companies release demos that at first glance appear to show agents cooperating with one another, but there are very few cases that rigorously and reproducibly demonstrate that such coordination is superior to a simple baseline (for example, a single agent or a rule-based script). Aftershock transforms coordination from “dramatic theater” into something “verifiable and measurable.”

The commander breaks down each mission into typed work units and assigns them according to role. Agents can choose to accept, reject, subdelegate, or escalate that work, providing a structured reason in each case. Limited resources (ambulances, rescue teams, fire engines, fuel, generators, etc.) compete through a typed proposal protocol. This protocol includes resource requests, task handoffs, escalations, and information sharing. Because the simulation engine atomically resolves all conflicts on every tick, coordination is implemented as clearly defined game mechanics rather than a “chat log.” It makes no unverifiable claims whatsoever. The “Decision Receipt” makes the rationale for every contentious decision traceable, and the “one-page Evidence Pack” directly links all key metrics (headline figures) to the specific source files in the repository.

Solution

AfterShock’s reliability is supported by its benchmark design. Disaster scenarios generated with the same seed value—completely identical disaster scenarios—are executed in the following four ways:

• Script-controlled bots (rule-based)
• Response by a single qwen3-max agent
• A flat swarm of qwen3.5-flash agents with no protocol (a non-hierarchical population)
• A structured society (hierarchical, with qwen3.5-flash agents as workers and qwen3.5-plus as the commander)

Because disasters use a fixed seed and are deterministic, all four conditions faceliterally the same emergency. Consequently, all differences in results are attributable to differences in coordination schemes, rather than to chance (being lucky or unlucky).

The evaluation metrics are the number of lives rescued, response latency, the number of failed missions, wall time, and cost. Rather than merely reporting averages, for the delta in each metric, the benchmark also provides bootstrap confidence intervals, paired sign-tests, and statistical power. When the effect is weak, it is explicitly labeled “suggestive,” avoiding unsupported causal claims. Furthermore, a “reproducibility badge” guarantees that the script-controlled engine is fully reproducible byte for byte. This is a benchmark harness that behaves like a careful scientist rather than a marketing page. For example, any rationale presented by an agent is always explicitly labeled “agent-stated” and is never presented as absolute ground truth.

Key Features

AfterShock’s live demo runs as a Docker Compose stack on Alibaba Cloud Elastic Compute Service (ECS). A FastAPI backend provides a React-based observability dashboard, while a Caddy HTTPS reverse proxy protects the frontend. All inference processing across the four benchmark conditions runs on Qwen models:

• Solo agent baseline: qwen3-max
• Flat swarm and structured society workers: qwen3.5-flash
• Structured society commander: qwen3.5-plus

By applying the same Qwen models and same-seed disaster scenarios to all four conditions, AfterShock enables fair, “apples-to-apples” comparisons that can answer questions such as “Does a hierarchical structure outperform a swarm?” and “Does either outperform a single powerful model?”

Alibaba Cloud Products Used

• Qwen Cloud
• Qwen Models
• ECS

About the Team

AfterShock was developed by Kenny Ademolu. He leads cyber architecture and engineering for building reliable AI-native systems in the fields of identity management, security, resilience, and agent evaluation. This project faithfully reflects his vision: rather than an agent society in which you “have no choice but to believe,” it realizes a society in which all numerical values, judgments, and supporting evidence are verifiable, reproducible, and statistically testable.

To learn more about the project: https://www.youtube.com/watch?v=TTM-iqX33qg
This project was developed as part of the global AI hackathon series hosted by Qwen Cloud.

This article is a translation of an Alibaba Cloud Blog article.