Planning for Agents

The text discusses planning and reasoning for agents, specifically focusing on Language Learning Models (LLMs). It explains that planning involves LLMs thinking about what actions to take in both short-term and long-term scenarios. Currently, developers use function calling to let LLMs choose immediate actions, but long-term planning is a more challenging task due to the growing context window and potential distractions for LLMs. The text also mentions that measuring the effectiveness of current models in planning and reasoning is difficult, with benchmarks like the Berkeley Function Calling Leaderboard being available. It suggests improving planning by ensuring LLMs have all required information and changing the cognitive architecture of applications. Two categories of cognitive architectures are discussed: general purpose and domain-specific. The text concludes that while LLM's planning and reasoning capabilities will likely improve, custom cognitive architectures will remain important for task-specific agents. It highlights LangGraph as a tool to facilitate building these custom architectures.