A new review maps how AI agents could help farming, and where they fall short

What the review looks at

As Biotech Bulletin (Shengwutong) reported, a review in npj Sustainable Agriculture examines how agentic AI could change farming. An AI agent here means software that takes in information, makes a decision and acts toward a goal, rather than just answering a question.

The authors say such systems could help farms work in complex, uncertain conditions, with possible gains in productivity, resource efficiency and food security. They also say the approach raises hard questions about robustness, explainability, ethics and fairness.

How these agent systems work

The review describes a loop: sense the situation, decide, act, then learn from the result. Many systems use a large language model, the same kind of AI behind chatbots, as the reasoning engine.

To reduce errors, some add retrieval, which means looking up outside documents before answering. Others use shared protocols, including the Model Context Protocol (MCP) for reaching outside knowledge and tools, and agent-to-agent protocols so several specialist agents can coordinate.

Examples and what they show

The review groups research into types, from tools that mainly hold farming knowledge to agents that make decisions and teams of agents. PhenoAssistant, which helps analyze plant traits, and Aleks, built for plant science, split a job among specialist roles. PAAS uses separate agents for tasks such as irrigation and leaf-disease checks.

The hard parts

The authors point to scarce, scattered and local farm knowledge, plus the tendency of language models to state false things confidently. A system trained in simulation can also fail in a real field, a gap often called simulation-to-reality.

Privacy and accountability are also open. If an agent's advice causes a loss, it is unclear whether the developer, the farmer or the platform is responsible.

Where it could go next

The review sees potential in measuring plant traits at scale, automated labs where AI plans and runs experiments, and robots guided by models that combine vision, language and action. It also mentions using AI to turn messy agricultural reports into structured databases.

This is a survey, not a new product. It has little connection to the Claude news in today's issue, but it shows how agent ideas are being tried in other fields.