The Rise of LLM-powered Agents: Reinventing Workflow Automation with Agentic AI

LLM-based applications have evolved from simple request-response systems, like ChatGPT, to tools that mimic reasoning, such as OpenAI's o1 and o3 models and Google's Gemini Thinking. Next on the scene is going to be the deployment of LLM-powered agents - autonomous entities that pursue goals, interact with their environment, and communicate with other systems to achieve goals. LLM-powered agents allow for autonomous and dynamic decision-making based on the execution context. This thesis builds upon Microsoft’s Magentic-One system, to explore its adaptability to more specific scenarios, assessing its potential as a foundation for agent-based applications. The work introduces an original agent-based workflow engine. The proposed system includes a set of agents, each specialized in a specific domain, collaborating to carry out the user tasks specified in natural language. This thesis also explores communication formats to increase the efficiency of agent interaction, enhancing scalability while ensuring reliability.