Evaluating AI agents isn't like testing traditional software where you can check if the output matches expected results. These agents perform complex tasks that often have multiple valid approaches, requiring them to understand context and follow specific rules while sometimes persuading or negotiating with humans. Researchers are tackling these challenges by examining fundamental capabilities that define an effective AI agent, each requiring its own specialized evaluation framework. The Berkeley Function Calling Leaderboard (BFCL) has pioneered a comprehensive framework for evaluating tool calling capabilities, which has evolved through multiple versions to address increasingly sophisticated aspects of function calling. Recent research has demonstrated that large language models can significantly enhance their problem-solving capabilities through self-reflection and evaluation, mimicking human metacognitive processes. Evaluation frameworks like Natural Persuasion in Open Discussions, Financial Manipulation Assessment, and Subtle Manipulation Through Language are being used to assess AI persuasion and manipulation capabilities, with results showing promising abilities while also highlighting the importance of robust safety measures. These efforts aim to support the development of AI systems that can engage effectively in legitimate persuasion while maintaining strong safeguards against harmful manipulation. By understanding these strengths and quirks, we're getting better at building AI systems that can truly complement human capabilities rather than just imitate them, shaping the future of human-AI collaboration.