PDMA-AI-Runtime: An Agent Orchestration Platform Based on the Perception-Decision-Memory-Action Model

This article introduces the production-grade AI agent orchestration platform PDMA-AI-Runtime, built on the Perception-Decision-Memory-Action (PDMA) model. It supports planning-driven orchestration, structured reasoning, MCP tool execution, and memory-perception iteration, providing a complete engineering solution for implementing autonomous workflows.

With the improvement of large language model (LLM) capabilities, AI agents are moving from proof-of-concept to actual deployment, but they face challenges such as coordinating multi-agent collaboration, managing long-term memory and context, and ensuring decision interpretability and controllability. Traditional LLM applications follow a request-response model, while AI agents require continuous state management, multi-step reasoning, and tool calling capabilities. This demands that the underlying platform has stronger orchestration capabilities to handle complex execution flows, error recovery, and state persistence.

The core innovation of PDMA-AI-Runtime is the PDMA execution model (Perception, Decision, Memory, Action), which draws on cognitive science research and decomposes the agent's cognitive process into four interconnected stages: The Perception stage collects information such as user input and system status and converts it into a structured internal representation; The Decision stage determines the next action based on perceived information and goals; The Memory stage manages long-term states (dialogue history, knowledge base, etc.) to maintain context coherence; The Action stage executes plans, calls tools, or generates responses.

PDMA-AI-Runtime provides key capabilities to support complex agent workflows: The planning-driven orchestration mechanism decomposes complex tasks into sub-task sequences and manages dependencies and execution order; Structured reasoning ensures model outputs follow predefined formats through prompt engineering and output validation; MCP tool execution capability supports agents to call external systems (APIs, databases, etc.) and provides full lifecycle management of tools; The memory-perception iteration loop stores dialogue history and retrieves relevant fragments, supporting short-term working memory and long-term knowledge memory.

The platform uses a centralized LLM Gateway to handle all model call requests, bringing benefits such as unified authentication and authorization, request rate limiting and quota management, dynamic model routing, and response cache optimization; The Gateway also undertakes observability responsibilities, recording metrics like latency and token consumption, supporting multi-model backend selection to balance quality, cost, and latency.

PDMA-AI-Runtime is suitable for complex multi-step reasoning and tool collaboration scenarios: In customer service, it supports intelligent customer service agents to understand complex problems and query knowledge bases; In R&D assistance, it builds development assistants to understand codebases and execute tests; In business process automation, it coordinates multiple systems to complete end-to-end processes (such as email requirement extraction, project status updates, etc.) to improve enterprise operational efficiency.

Compared to frameworks like LangChain and LlamaIndex, PDMA-AI-Runtime focuses more on reliability and operability in production environments, providing comprehensive error handling, state persistence, and monitoring capabilities; The explicit phase division of the PDMA model makes system behavior more predictable and debuggable, allowing developers to clearly see the input and output of each cognitive phase for easy diagnosis and optimization.

PDMA-AI-Runtime is an open-source project with a modular architecture design, where components interact through clear interfaces to improve maintainability and community contribution; In the future, with the development of multimodal large models and embodied intelligence, the platform needs to focus on the scalability of the PDMA model and support for new modalities such as vision and voice.