A Georgia Tech Gardening AI Splits Its Advice Among Three Specialists, and Its Mistakes Teach as Much as Its Answers

Ask a general-purpose chatbot when to plant tomatoes and you will usually get an answer. Whether it is the right answer for a backyard in Gilbert, a balcony in Atlanta or a rented lot in Honolulu depends on details the chatbot may never ask about.

A research team at the Georgia Institute of Technology set out to design around that problem. Their system, called CultivAgents, splits gardening advice among three specialized AI “agents,” each built to handle one part of the question, and then lets them take turns answering. In a small study described in a paper posted to the research archive arXiv in May 2026, community gardeners who tried it came away more confident and more willing to act on AI advice. Expert reviewers, meanwhile, caught it making the kinds of mistakes that can cost a gardener a season.

The project was led by Yiyang “Diana” Wang of Georgia Tech, working with Moeiini Reilly, Britney Johnson, Kefei Yan and Josiah Hester of Georgia Tech and Alex Cabral of the Massachusetts Institute of Technology.

Why gardening advice is hard to automate

The researchers start from a familiar problem. Citing earlier research, they note that roughly half of new food gardeners quit after their first season. People who move to a new region face their own version of the same struggle: unfamiliar climate, soil and pests, and less access to the plants and know-how they grew up with.

Good advice, the team argues, has to account for plant biology, local climate, soil, season, pests, available materials and the gardener’s own cultural knowledge. A single chatbot tends to blend all of that into one generic reply. As the paper puts it, ungrounded advice “can lead to wasted effort, failed crops, inappropriate chemical use, or unsafe plant-care decisions.”

Three agents, one conversation

CultivAgents opens each session by asking the user four things: gardening experience level, location, the current month and cultural background. That profile is attached to the instructions for every agent.

From there, three agents share the conversation. An Experience Agent adjusts complexity and tone, offering step-by-step help to beginners and more technical notes on cultivars, soil amendments and pest management to experienced growers. An Environmental Agent gives location-specific, seasonal guidance, referencing local soil conditions, frost dates and extension resources. A Cultural and Ethnobotanical Agent shares historical and medicinal plant knowledge from different cultures, and is designed to separate traditional uses from scientifically validated ones.

A separate selector decides who speaks next, routing technical questions to the Experience Agent, seasonal or location questions to the Environmental Agent and cultural questions to the Ethnobotanical Agent. The same agent cannot speak twice in a row, which pushes the system to offer different perspectives, and each round ends after three messages. In the chat window, the agents are color-coded green, blue and orange so users can tell them apart. The team built it with Microsoft’s open-source AutoGen framework for multi-agent AI systems.

What the gardeners said

The evaluation had three phases. Three domain experts in Hawaii and Georgia, people who design and use agricultural technology in food sovereignty and education work, reviewed the system first. Seven human-computer interaction researchers at Georgia Tech tested it next. Finally, five university students involved in campus community gardening used it in a workshop and filled out surveys before and after.

On five-point scales, the five gardeners’ average confidence rose from 3.00 to 3.60, motivation rose from 4.00 to 4.40, and trust in acting on AI gardening advice rose from 3.20 to 4.00. They gave the system an average usability score of 84.5 on a standard 100-point questionnaire, well above the commonly used benchmark of 68.

The Environmental Agent was the favorite. It scored 4.80 out of 5 for location relevance, and three of the five gardeners picked it as the most helpful agent. The group also agreed, with an average of 4.60, that multiple specialized agents were better than a single general chatbot. One participant said they would feel “less confident” relying on one agent to explain everything. Another said the cultural agent “expands my knowledge past just the question I asked.”

The gardeners also asked for things the system did not yet do: pictures and visual aids, planting schedules, more information on invasive plants, and follow-up prompts that point toward a “next logical question.”

The numbers come with a large caveat. Five people is a very small sample, and they were students already active in a community garden, so the results show how a motivated group reacted rather than how the public would.

The mistakes

The most useful part of the paper for everyday gardeners may be what the experts found. They said the system’s answers were often limited by “generalized LLM knowledge,” meaning the broad but shallow knowledge built into large language models.

Some errors were about place. Experts flagged suggestions of plants that were poorly matched to wet soil, and species that were inappropriate or invasive for the user’s location, including mangroves recommended for Hawaii. They also noted that real constraints such as apartment rules and dense urban shade take more than general climate data to understand.

Others were about culture. When one expert identified as Ilokano, a people of the northern Philippines, the system suggested ube and used the Tagalog word “gabi” for taro instead of the Ilokano word “aba.” The researchers describe this as “cultural flattening”: answering with broad associations instead of attention to the person’s specific community.

The team also found that agents answering independently could feel repetitive or even contradict each other. Their recommendation is coordinated expertise, with shared memory, clearer roles and checks for contradictions, rather than simply more voices.

What this means for a desert garden

CultivAgents was a research prototype, and the paper links to a demonstration deployment rather than a finished product. But its design offers a practical checklist for anyone in San Tan Valley or the wider East Valley who already asks an AI tool for garden help.

Start by giving the tool the same four facts CultivAgents collects: how experienced you are, where you live, what month it is, and any plants or traditions that matter to you. “Low desert, Pinal County, late October, second-year gardener” will get a different answer than “when should I plant carrots?”

Then check the place-specific parts. The experts’ mangrove example is a reminder that a confident AI answer can still suggest a plant that does not belong. In the Sonoran Desert, local sources such as University of Arizona Cooperative Extension publications, county Master Gardener programs and native plant groups are the right places to confirm planting dates, water needs and whether a plant is considered invasive here.

Finally, take the paper’s own advice about relationships. The researchers write that gardening systems should help connect users to local gardeners, elders, extension services, seed libraries and community organizations when knowledge is place-based. An AI tool can help a new gardener find the right question. The neighbors, libraries and extension volunteers who know this soil are still the best people to help answer it.