Plan-and-Execute Agent

LangChain has released a new 'Plan-and-Execute' agent executor, distinct from its existing ReAct-based 'Action' agent. Inspired by BabyAGI and the Plan-and-Solve paper, this approach excels at establishing complex long-term plans.

The existing Action agent follows an iterative structure where it receives user input, selects and executes tools, and then determines the next step based on the results. However, as user goals become more complex and reliability in production environments becomes critical, the accumulation of past history and tool usage guidelines has led to a rapid increase in prompt size.

The Plan-and-Execute agent addresses these issues by separating high-level planning from short-term execution. The Planner leverages the reasoning capabilities of language models to handle ambiguity and establish step-by-step plans, while the Executor determines and executes the appropriate tools for each step. Although this structure has the downside of increasing the number of language model calls, it is advantageous for handling complex requests and ensuring reliability.

Currently, this feature is implemented in Python and TypeScript and is provided within the experimental module due to anticipated rapid changes.