AI Agronomist Agent (CARL)
## key insight
The interface was the easy part. Co-defining the AI system underneath it — which tools the agent uses, how to communicate uncertainty to domain experts, what MCP integration standards look like — that was the actual design challenge.
## the challenge
Field agronomists and seed sales reps are PhD-level domain experts making consequential on-farm decisions in real time. Every answer has financial and agronomic stakes. The AI had to earn trust from users who are professionally skeptical of software that oversimplifies — and it had to communicate uncertainty honestly.
## architecture & process
┌─Conversational UX + decision flows· natural-language interface for agronomists
├─AI system architecture co-design· tools selection, system prompts, agentic orchestration
├─MCP server standards· framework for internal teams building on CARL
├─Agronomic data sources· research library, yield data, Pioneer product info
├─Uncertainty communication patterns· designed for domain experts with real consequences
└─AI design system· tokens, components, interaction patterns — 3 squads
## key decisions
- Went beyond UX — co-defined the AI system architecture (tools, system prompt design, voice and tone, agentic orchestration). Unusual for a designer. Critical for getting the product right.
- Defined standards for how internal teams build, test, and integrate MCP servers with CARL — establishing a replicable pattern across the org.
- Designed for latency: agricultural decisions can't wait on a spinner. Feedback patterns make response time feel purposeful, not broken.
- Co-developed the 6-week iterative sprint framework — validation with real users before engineering commits. Adopted as standard across all squads. Peer squads took twice as long.
- Mentored 2 designers to full surface ownership. Influenced 6+ through structured design reviews.
## results
5%→75%
AI adoption by agencies
+50%
data sharing, agencies ↔ field reps
½
peer teams' discovery time
2+6
designers mentored + influenced
## tools used
FigmaMCPRAGLangchainClaude CodeScalable DesignData Governance
## screens
## field notes
“Gabriel is the rare designer who treats the prototype as the spec. He'll ship a working flow before most teams have finished arguing about it.”