Designing and Implementing a Configurable Engine for Dynamic LLM Agent Orchestration Across Multiple Platforms

Over the past year, advances in large language models have increased demand for the technology in production environments. Many industries are now building AI chat assistants that understand natural language, retrieve information and perform specific tasks. However, building such a system is complex. It requires developers to design the entire pipeline to manage the conversation and integrate external tools to carry out tasks. Most of this logic is usually written directly in source code. This ties the dialogue to the implementation, so every change to a conversation means editing and retesting the code, and the resulting assistants become fragile and hard to adapt to new user needs. This thesis presents a new approach to developing such agentic systems. Instead of coding the whole process directly in the source code, the behavior of the system is described declaratively in human-readable JSON configuration files. An Engine, built on top of the LangGraph and LangChain frameworks, translates these configurations into a working agent pipeline. It reads the graph definition, imports and links the required handler functions dynamically, and constructs the complete conversational execution graph at runtime. The Engine also supports conditional routing, nested subgraphs for modular workflows, and human-in-the-loop interactions for handling complex multi-turn conversations. To extend agent capabilities without hardcoding external dependencies, the system integrates the Model Context Protocol (MCP), which standardizes how tools are discovered and executed. In addition, a channel abstraction layer decouples graph execution from the client interface, so the same conversational assistant can be deployed on different platforms, such as terminal consoles and messaging applications like Telegram. To validate the proposed architecture, a complete multi-service application, AutoData AI, was implemented for an automotive dealership scenario, showing that complex domain-specific workflows can be orchestrated and maintained purely through declarative configuration.