An Agentic Ai Framework For Autonomous Reasoning, Tool Selection, And Task Execution Using Large Language Models

Abstract

The increasing complexity of AI-assisted workflows requires autonomous systems capable of reasoning, selecting tools, and executing multi-step tasks beyond conventional text generation. Existing large language model (LLM) assistants often depend on static prompting and exhibit limitations in planning, tool selection, execution monitoring, and recovery from intermediate failures. This study proposes an Agentic AI Framework (AAIF) that integrates task interpretation, goal decomposition, autonomous reasoning, dynamic tool selection, task execution, execution-state monitoring, error recovery, and final-response generation. The LLM acts as the central reasoning and orchestration engine, while external software capabilities are organized through a structured tool registry. AAIF is computationally evaluated on representative tasks with simple, moderate, and complex workflows and compared with Direct LLM, Fixed-Tool LLM, and Planning-Only Agent approaches. Performance is assessed using task success rate, tool-selection accuracy, planning success rate, execution efficiency, average execution time, and recovery success rate. The proposed framework achieves 94.2% task success, 96.1% tool-selection accuracy, 93.4% planning success, 91.8% execution efficiency, 2.84 s average execution time, and 92.6% recovery success, demonstrating improved reliability for autonomous LLM-driven task execution.