Should AI Agents Act Through Code? | CodeAct Explained

What if an AI agent could write a small program instead of making one tool call at a time? In episode 04 of our CMU AI Agents reading series, we explore Executable Code Actions Elicit Better LLM Agents, the paper behind CodeAct. We explain how executable Python lets agents combine tools, reuse intermediate results, and respond to execution feedback across multiple turns. We walk through the comparison with JSON and text actions, the M3 ToolEval benchmark, and the CodeActInstruct training data used to build CodeActAgent. We also examine where the approach falls short—and why fewer interaction turns do not automatically mean lower cost, faster execution, or guaranteed correctness. Paper: Xingyao Wang et al., Executable Code Actions Elicit Better LLM Agents (ICML 2024). https://arxiv.org/abs/2402.01030 CMU AI Agents course: https://www.cmu-agents.com/ Beyond the Abstract is an independent paper reading group following the course, not affiliated with CMU or the paper's authors. This video is AI-assisted. Paper visuals: Figure 1, Table 3, and Table 5 from Wang et al., arXiv:2402.01030v4, cropped for readability. Results reflect the models and benchmarks studied in 2024. When would you choose constrained tool calls, and when would you let an agent write a program? Share your ideas and corrections in the comments. Stay curious. Keep learning. #AIAgents #CodeAct #ToolUse #LLM #BeyondTheAbstract