The Future of AI Agents in Health and Nutrition

By 2035, over half the global population is expected to be overweight, costing the world economy an estimated $4 trillion. AI advancements, particularly with LLMs like GPT-4o, Claude 3, and Gemini 1.5 Pro, are paving the way for intelligent, domain-specific agents. These agents promise to transform healthcare from treatment to prevention, offering personalized and proactive guidance.

Why Health AI Agents Matter

Imagine a billion people with access to an always-on AI nutrition coach. Diets would improve, chronic disease rates would fall, and healthcare would shift focus. However, today’s agents are mostly generic copilots or chatbots. Truly useful, vertical-specific agents that operate within a domain and deliver consistent, meaningful action are still rare.

Designing AI Agents from First Principles

Building effective agents requires rethinking traditional apps from the ground up. For example, Welling, a health-focused agent, didn’t just upgrade existing calorie-tracking apps like MyFitnessPal with AI. Instead, it reimagined the entire experience, asking: If language models were the default UI, how would a nutrition coach work?

The UI of an AI Agent

While chat is the most intuitive interface, pure chat interfaces like WhatsApp or Telegram are limiting. Modern agentic UIs are evolving into hybrid interfaces—chat-driven bases layered with dashboards and dynamic components. For instance, ChatGPT now includes commerce and shopping cards in its chat. This new design paradigm comes with challenges:

• How to separate input (e.g., chat) from output (e.g., dashboards) without confusing the user?
• Should the agent decide or the user? For example, should water intake be logged automatically or manually?
• Should all input data go through the agent, or should some be manual?

Solving the Messy Problem Space of Language

LLMs excel with language, but humans are often vague. Had a bit of curry rice isn’t a database entry—it’s a probabilistic phrase. Vertical agents need to build their own logic layer to handle domain-specific use cases. Welling, for instance, built a food parsing engine using context-aware prompts, search algorithms, and user feedback to ensure reliable calorie and macro estimations.

Deterministic Outputs from Probabilistic Systems

Foundational models are probabilistic, but users in domains like health, legal, or finance expect consistency. Welling ensures the same input (e.g., ham, egg, tomato sandwich) yields the same calorie estimate every time. This requires sophisticated guardrails and deterministic logic layers.

Memory and Context: What to Store

With expanding context windows (e.g., Gemini 2.0’s 1M tokens), the challenge isn’t storage capacity but deciding what to store and how to process it. For example, when a user asks about breakfast, should Welling include 30 days of eating history or an abstracted layer? Tools like and help manage long-term memory and data access.