A Benchmark Framework for Evaluating Agentic AI Systems in Real-World Tasks

The rapid proliferation of large language model (LLM) based autonomous agents has created an urgent need for standardised evaluation frameworks capable of capturing multi-step, tool-interactive behaviour. We introduce AgentEval, a multi-dimensional benchmark framework evaluating LLM-based agents across five independent dimensions: Task Success, Efficiency, Tool Usage, Reasoning Quality, and Robustness. Each dimension is operationalised through two sub-metrics measured via automated procedures. Experiments comparing LLaMA-3.1-8B-Instant (Groq API) and TinyLLaMA-1.1B (Ollama, local) on a 20-task benchmark show overall completion rates of 0.900 vs. 0.600, with the largest performance gap in Robustness (1.000 vs. 0.400) and equal performance on Coding tasks (1.000 vs. 1.000). All code, tasks, and results are released as open-source resources.