How Moda Built a Production-Grade AI Design Agent Using Deep Agents

Moda is an AI-native design platform for non-experts that provides a Cursor-style AI sidebar, allowing users to directly generate and edit designs on a 2D vector canvas.

At the core of the system is a multi-agent architecture based on Deep Agents, with observability secured through LangSmith to support rapid iteration and stable deployment.

Moda's agent system performs three roles as follows:

• Design Agent: Handles real-time design generation and iteration on the canvas.
• Research Agent: Fetches structured content from external sources and saves it to the user's file system.
• Brand Kit Agent: Learns brand assets (colors, fonts, logos, etc.) to maintain design consistency.

Unlike HTML/CSS, visual design lacks layout abstraction and uses XML-based absolute coordinates (XY), making it difficult for LLMs to handle. To address this, Moda developed a custom DSL (Domain Specific Language) that makes structure easier for LLMs to understand instead of raw coordinates, reducing token cost and improving quality.

Every request first goes through a Triage stage using the lightweight, inexpensive Haiku model. Here, the output format is classified and relevant Skills (such as design guidelines) are pre-loaded to maximize the efficiency of the main agent.