Moda Deploys Multi-Agent AI System for Professional Design Automation

Moda has revealed the technical architecture behind its AI-native design platform, showcasing a three-agent system built on LangChain’s Deep Agents framework that lets marketers, founders, and small business owners produce professional presentations and marketing materials without design expertise.

The platform—positioned as a Canva alternative with Cursor-style AI assistance—addresses a fundamental problem in AI-generated design: LLMs struggle with visual layouts because they’re terrible at reasoning about pixel coordinates.

Why PowerPoint’s DNA Breaks AI Design

Most design tools rely on formats like PowerPoint’s XML spec, which uses absolute XY coordinates to position elements. “LLMs are not good at math,” said Ravi Parikh, highlighting why AI-generated decks typically look generic and poorly composed.

Moda’s solution? A proprietary domain-specific language (DSL) that gives the AI layout abstractions rather than raw numerical coordinates—similar to how Flexbox and CSS grid make web development accessible to language models. The company isn’t sharing specifics, but the approach reportedly cuts token costs while improving output quality.

Three Agents, One Canvas

The system runs three specialized agents:

Design Agent handles real-time creation and iteration through the AI sidebar. This runs on a custom LangGraph loop, though migration to Deep Agents is under evaluation.

Research Agent pulls structured content from external sources like company websites, storing it in a per-user file system. Already running on Deep Agents.

Brand Kit Agent ingests colors, fonts, logos, and brand voice from websites, uploaded guidelines, or existing decks. Also Deep Agents-powered.

All three share a common architecture: lightweight triage using fast Haiku models, dynamic context loading, and full observability through LangSmith tracing.

The Context Engineering That Actually Matters

Moda’s triage system classifies each request by output format—slide deck, PDF, LinkedIn carousel—then pre-loads relevant “skills,” which are Markdown documents containing design best practices and format-specific instructions. Prompt caching breakpoints sit after the system prompt and skills block, keeping frequently-used context cached while allowing dynamic injection.

The Design Agent maintains 12-15 core tools in context, with roughly 30 additional tools available on demand through a RequestToolActivation call. Each extra tool costs 50-300 tokens and breaks prompt caching, but the data shows most requests don’t need them.

For large projects—say, a 20-slide deck—the system dynamically manages context, providing high-level summaries and letting the agent pull details as needed rather than loading everything upfront.

Collaboration Over Generation

What separates Moda from typical AI design tools is the interaction model. Instead of generate-and-replace, the AI works directly on a fully editable 2D vector canvas. Every element remains selectable, movable, and styleable. The relationship shifts from “accept or reject” to genuine back-and-forth refinement.

The platform has reportedly found early traction with B2B companies doing enterprise sales—teams that need polished pitch decks fast but want control over the final product. Integration with Microsoft 365 workflows adds enterprise appeal.

What’s Coming

Moda’s roadmap includes completing the Deep Agents migration for the Design Agent, activating memory primitives already in place, and expanding brand context support for multi-team enterprise customers. Formal evaluation systems are planned but not yet implemented—for now, LangSmith traces serve as the primary feedback mechanism for catching regressions.

The technical disclosure offers a useful blueprint for teams building production AI agents: custom context representations beat raw data formats, dynamic tool loading outperforms kitchen-sink approaches, and observability isn’t optional when you’re shipping to real users.

Image source: Shutterstock