Evaluating a ReAct-Based Agent for Agricultural Planning

Karanbir Singh, Jaskirat Singh

Abstract:

Planning is a hallmark of intelligence, enabling both humans
and artificial agents to navigate complex environments, adapt
to dynamic conditions, and achieve intricate goals.
Significant emphasis has been placed on enhancing the
planning capabilities of large language models (LLMs) using
agents. ReAct, an agentic framework, introduces a paradigm
that combines reasoning traces and actionable steps in an
interleaved, iterative process, allowing LLMs to adaptively
plan based on real-time feedback. This paper investigates the
application of ReAct-based agents in agricultural planning,
focusing on gardening activity planning as a testbed and
introduces the Gardening Planner, an agent that integrates
reasoning capabilities with tools such as a Retrieval-
Augmented Generation (RAG) system, a weather forecast
API, and a dynamic web search tool. The agent leverages
these components to create personalized, context-aware crop
plans for the current season. Our agent can provide effective
crop plans only 50% of the time. The findings reveal key
limitations of ReAct-based agents in dynamic, real-world
settings such as agriculture, emphasizing their reliance on
retrieval and summarization over genuine planning. This
study also contributes to the broader discourse on planning
and reasoning in AI by highlighting the challenges of
deploying agentic frameworks in practical applications.