Phleet: An Open-Source Autonomous Multi-Agent AI Platform Based on Temporal Workflow

Phleet is an open-source autonomous multi-agent AI platform deployed via Docker containerization. It supports agents powered by Claude or Codex, orchestrates them through Temporal workflows, and allows coordination and control via Telegram. Addressing the orchestration challenges of multi-agent systems, it combines Docker containerization, Temporal workflow engine, and instant messaging coordination to provide a complete solution for building complex agent applications.

With the improvement of large language model capabilities, multi-agent application architectures have evolved rapidly, but they face three major challenges: coordination (sequential execution, dependency management, shared state management), reliability (fault tolerance and recovery), and observability (tracking of running states). Traditional task queues or messaging systems are difficult to meet the requirements, so Phleet chooses Temporal as the workflow engine to address these issues.

Advantages of Temporal Workflow

• Persistence and Fault Tolerance: Workflow states are persisted, automatic retries on failure, process crash recovery—suitable for long-running multi-agent workflows.
• Programming Model: Supports complex logic like conditional branches, loops, parallel execution—flexible definition of agent collaboration patterns.
• Observability: Provides tools for workflow history and execution status, facilitating problem localization.

Phleet Architecture

• Dockerized Agents: Each agent is encapsulated as an independent container, with strong isolation, portability, and easy scalability—focus on business logic without worrying about infrastructure.
• Workflow Definition: Declaratively define multi-agent collaboration processes (e.g., multi-step chaining for content creation) via Temporal.
• Telegram Coordination Interface: Users can start workflows, check status, and receive notifications via Telegram, leveraging its wide usage and Bot API features.

Phleet适用于多种场景：

1. Automated Workflows: Chaining multi-stage agents (e.g., intent recognition, information query in customer service).
2. Complex Task Decomposition: Splitting tasks into specialized agents (e.g., literature retrieval, report writing for research tasks).
3. Human-Agent Collaboration: Inserting manual review nodes in workflows, combining AI automation with human judgment.
4. Long-Running Processes: Leveraging Temporal's persistence to handle long-cycle projects (e.g., market research).

• Agent Communication: Transfer data and trigger agents via Temporal workflow context, avoiding direct coupling.
• State Management: Temporal automatically manages workflow states; cross-workflow shared states can use query functions or external storage.
• Error Handling: Use Temporal's retry policies, compensation transactions, and timeout control to ensure reliability.
• Scalability: Docker containerization supports horizontal scaling; Temporal cluster deployment provides high availability and load balancing.

• Learning Curve: Temporal has complex concepts; need to optimize documentation and examples to reduce the learning barrier.
• Resource Overhead: Docker and Temporal bring certain overhead; need to optimize resource usage or provide lightweight deployment options.
• Ecosystem Integration: Currently only supports Claude/Codex and Telegram; need to expand support for other LLMs and communication platforms.

Phleet demonstrates a way to build reliable and scalable multi-agent systems, providing an open-source solution for complex AI applications. Although not the simplest option, it is highly valuable for scenarios requiring complex workflows, high reliability, and human-agent collaboration. As multi-agent systems become more popular, such infrastructure projects will lower the development threshold, allowing developers to focus more on business logic.