Peter Zhang
Mar 24, 2026 17:27
Moda’s Deep Brokers-powered platform allows non-designers to create production-grade visuals by way of a three-agent structure with customized DSL for structure reasoning.
Moda has revealed the technical structure behind its AI-native design platform, showcasing a three-agent system constructed on LangChain’s Deep Brokers framework that lets entrepreneurs, founders, and small enterprise homeowners produce skilled shows and advertising and marketing supplies with out design experience.
The platform—positioned as a Canva different with Cursor-style AI help—addresses a elementary downside in AI-generated design: LLMs battle with visible layouts as a result of they’re horrible at reasoning about pixel coordinates.
Why PowerPoint’s DNA Breaks AI Design
Most design instruments depend on codecs like PowerPoint’s XML spec, which makes use of absolute XY coordinates to place parts. “LLMs are usually not good at math,” mentioned Ravi Parikh, highlighting why AI-generated decks usually look generic and poorly composed.
Moda’s answer? A proprietary domain-specific language (DSL) that offers the AI structure abstractions quite than uncooked numerical coordinates—just like how Flexbox and CSS grid make internet improvement accessible to language fashions. The corporate is not sharing specifics, however the method reportedly cuts token prices whereas enhancing output high quality.
Three Brokers, One Canvas
The system runs three specialised brokers:
Design Agent handles real-time creation and iteration by way of the AI sidebar. This runs on a customized LangGraph loop, although migration to Deep Brokers is beneath analysis.
Analysis Agent pulls structured content material from exterior sources like firm web sites, storing it in a per-user file system. Already working on Deep Brokers.
Model Package Agent ingests colours, fonts, logos, and model voice from web sites, uploaded tips, or current decks. Additionally Deep Brokers-powered.
All three share a standard structure: light-weight triage utilizing quick Haiku fashions, dynamic context loading, and full observability by way of LangSmith tracing.
The Context Engineering That Really Issues
Moda’s triage system classifies every request by output format—slide deck, PDF, LinkedIn carousel—then pre-loads related “abilities,” that are Markdown paperwork containing design greatest practices and format-specific directions. Immediate caching breakpoints sit after the system immediate and abilities block, protecting frequently-used context cached whereas permitting dynamic injection.
The Design Agent maintains 12-15 core instruments in context, with roughly 30 extra instruments accessible on demand by way of a RequestToolActivation name. Every further instrument prices 50-300 tokens and breaks immediate caching, however the information reveals most requests do not want them.
For giant tasks—say, a 20-slide deck—the system dynamically manages context, offering high-level summaries and letting the agent pull particulars as wanted quite than loading the whole lot upfront.
Collaboration Over Era
What separates Moda from typical AI design instruments is the interplay mannequin. As an alternative of generate-and-replace, the AI works instantly on a completely editable 2D vector canvas. Each aspect stays selectable, movable, and styleable. The connection shifts from “settle for or reject” to real back-and-forth refinement.
The platform has reportedly discovered early traction with B2B firms doing enterprise gross sales—groups that want polished pitch decks quick however need management over the ultimate product. Integration with Microsoft 365 workflows provides enterprise enchantment.
What’s Coming
Moda’s roadmap contains finishing the Deep Brokers migration for the Design Agent, activating reminiscence primitives already in place, and increasing model context help for multi-team enterprise prospects. Formal analysis programs are deliberate however not but applied—for now, LangSmith traces function the first suggestions mechanism for catching regressions.
The technical disclosure affords a helpful blueprint for groups constructing manufacturing AI brokers: customized context representations beat uncooked information codecs, dynamic instrument loading outperforms kitchen-sink approaches, and observability is not elective if you’re transport to actual customers.
Picture supply: Shutterstock
