AgentFlow

Resilient Adaptive Cloud-Edge Framework for Multi-Agent Coordination

AgentFlow is a Multi-Agent System (MAS)-based framework for scalable, fault-tolerant coordination across heterogeneous cloud-edge infrastructures. Designed for real-time and mission-critical applications, AgentFlow enables decentralized, dynamic service orchestration using programmable logistics objects and abstract agent interfaces.

• Decentralized Decision Making
  Supports agent-driven coordination without central servers using lightweight consensus.

• Programmable Logistics Objects
  Request/Response Logistics allow precise, selective communication, minimizing unnecessary network traffic.

• Dynamic Service Election
  Agents elect optimal services at runtime based on real-time load and responsiveness.

• Many-to-Many Coordination Model
  Enables scalable task distribution and autonomous recovery in complex cloud-edge environments.

• Resilience and Fault Tolerance
  Supports agent-level fault containment and task reassignment under node failures.

• Modular Architecture
  Holonic agent structure with loosely coupled communication layers (e.g., MQTT, DDS).

+----------------------------+
|     Orchestration Layer    | <-- Load balancing, no central control
+----------------------------+
|         Agent Layer        | <-- Holonic agents: Perception, Decision, Action
+----------------------------+
|     Communication Layer    | <-- MQTT / DDS pub-sub abstraction
+----------------------------+

AgentFlow employs an event-driven publish-subscribe pattern with three logistics mechanisms:

• Selective Request-Response: Each client gets a unique topic to prevent message broadcasting.
• Dynamic Election: Tasks are dynamically assigned to the least-loaded agent.
• Composite Coordination: Coordinators manage agent clusters for many-to-many interactions.

Service Agent Selection:

Communication Mapping:

Tested using a swarm of 50–500 autonomous mobile robots (AMRs):

• Smart warehouses and AMR fleets
• Industrial IoT and edge robotics
• Intelligent grid and healthcare logistics
• Programmable, real-time distributed systems

AgentFlow/
├── src/agentflow/
│          ├── broker/    # MQTT/DDS brokers
│          ├── core/    # Holonic agent definitions
│          └── logistics/    # Request/response/election logistics
├── unittest/    # AgentFlow unittests
└── README.md

For questions or collaboration inquiries, contact:

• Ming Fang Shiu – 108582003@cc.ncu.edu.tw
• Prof. Ching Han Chen – pierre@csie.ncu.edu.tw