How Moda Built a Design Agent That Doesnt Produce the Same Generic Slides Everyone Else Gets
The pitch is familiar by now: AI generates your slides, you tweak them, done. The output is also familiar: same layouts, same stock-adjacent photography, same fonts someone picked because they were already installed. It looks like everyone elses deck because in a meaningful sense, it is.
Moda, which launched publicly today with a $7.5 million seed round led by General Catalyst (with participation from Pear VC, WndrCo, and angels from Dropbox, Stripe, Segment, and Scale AI), is trying to solve a different part of the problem. Rather than generating static images or applying templates, Moda gives every team a multi-agent system that learns its visual language and produces fully editable designs on a collaborative canvas. The company claims the result actually looks like your company, not like a prompt-engineered hallucination of professional.
The founding team is worth noting. CEO Anvisha Pai co-founded Dover (AI recruiting, $23M raised from Y Combinator, Founders Fund, and Tiger Global) and worked at Dropbox pre-IPO. COO Ravi Parikh co-founded Heap (acquired by Contentsquare for $960M valuation) and Airplane (acquired by Airtable). CTO John Holliman was employee #1 at Dover and scaled infrastructure at Expanse before its acquisition by Palo Alto Networks. These are operators who have shipped to scale, not first-time founders running up a seed round on a landing page.
The product story is where it gets interesting for the agent infrastructure beat.
Three agents, one canvas, no templates
Modas backend runs three distinct agents, coordinated through LangChains stack. The Design Agent handles real-time creation and iteration on the canvas — its a custom LangGraph loop that predates the teams investment in Deep Agents and is currently being evaluated for migration. The Research Agent and Brand Kit Agent both run on Deep Agents, LangChains open-source agent harness released March 23, 2026 under an MIT license.
The architecture is worth understanding because it says something about where LangChain is positioning itself. Deep Agents is not a novel runtime — its a LangGraph graph with opinionated defaults: a detailed system prompt, a no-op Todo list planning tool (the same pattern Claude Code uses, which is exactly the point), sub-agent spawning via isolated context windows, and file system access for shared workspace. The GitHub repo describes it as "a working agent immediately, customize what you need" — a harness for teams that want production-grade scaffolding without wiring it from scratch.
Modas use of it is a signal: this is exactly the kind of vertical, production workload LangChain is betting will migrate from "build your own agent loop" to "install a harness and configure it." Whether that bet pays off depends on whether the MIT-licensed toolkit holds up under real workloads outside LangChains own demo cases.
The design-specific engineering
Whats genuinely novel in Modas implementation is not the multi-agent choreography — thats a LangGraph workflow pattern thats been in production at enterprise AI shops for 18 months. Its the context engineering specific to visual design.
The core problem Moda identified: LLMs are bad at math, and PowerPoints XML spec is a verbose collection of XY coordinates that models reason about poorly. Raw canvas state is token-expensive and coordinate-heavy — not a natural representation for how models think about layout and structure. Modas solution was to build a custom context representation layer, a DSL that gives the agent a cleaner, more compact view of whats on the canvas. The specifics are proprietary, but the principle is the same one that makes LLMs effective at web development: give the model layout abstractions it can reason about, not raw numerical coordinates.
Ravi Parikh, Modas COO, put it plainly in the LangChain case study: LLMs are not good at math. PowerPoints XML spec has a bunch of XY coordinates — thats a fine representation of the data, but its not a great way for an LLM to describe where it wants things to live.
The triage step is also worth noting. Every request passes through a lightweight Haiku-based classifier that determines output format (slide deck, PDF, LinkedIn carousel, logo) and pre-loads the relevant skills — Markdown documents containing design best practices, format guidelines, and creative instructions. Skills are injected as human messages with prompt caching breakpoints placed after the system prompt and after the skills block, keeping the system prompt fixed and cached while allowing dynamic context injection per request.
The tool loading strategy reflects real tradeoffs. The Design Agent runs with 12-15 core tools in context at all times. An additional ~30 tools are available on demand via RequestToolActivation when the agent recognizes a specialized need, like parsing an uploaded PowerPoint. Each additional tool costs 50-300 tokens in the prefix and breaks prompt caching. The math works out because most requests dont need the extra tools — but its a reminder that token economics are a first-class design constraint in production agent systems, not an afterthought.
What LangChains involvement means
The case study was published on the LangChain blog, which means this is partly a product demo. Thats worth saying plainly. LangChain has a business interest in showing Deep Agents running real production workloads, and the $7.5M Moda raise on the same day as the Deep Agents 0.5.0 release (March 23, 2026) suggests coordination, not coincidence.
Thats not a criticism — the value of a platform is proven by what builders ship with it. But it means the technical claims in the case study should be read as Modas experience, not independently verified benchmarks. The piece notes that Moda doesnt yet run formal evaluations; LangSmith traces currently serve as the primary feedback loop for catching regressions. Formal evals are on the roadmap. For a system making claims about output quality and brand alignment, thats a gap worth watching.
The LangSmith observability layer is where the articles claims are most credible. Because all three agents are traced, Modas team has per-node token costs, cache hit rates, and tool call failures visible in production. Parikh keeps the trace view open while actively developing. That kind of operational transparency is the right answer to "how do you know its working" — its just not the same as third-party evaluation.
What this means for builders
The broader pattern here is LangChains push from infrastructure into packaged agent products. Deep Agents sits between "heres LangGraph, wire it yourself" and "heres a vertical SaaS with agents inside." The MIT license means it can be installed and modified freely; the opinionated defaults mean a team can go from pip install to a running agent in minutes rather than days.
For builders evaluating agent frameworks, Modas choices are instructive. The triage, skills, main loop pattern is a clean separation of concerns. The DSL approach to context representation is a reminder that the interface between your agent and your domain is a first-class engineering problem, not a prompt engineering afterthought. And the tool caching economics are a useful data point: keeping context lean wins more often than it loses.
The honest question is whether brand memory and taste are durable differentiators or marketing language. Moda says its system indexes a companys website, Google Drive, past slide decks, and existing assets to build a visual model that improves with every design. Thats a compelling claim. Its also the same claim that every AI-native design tool has made for two years, and the output gap between promise and reality has been consistently wider than the marketing suggests.
Modas specific bet is that the combination of multi-agent orchestration, editable canvas output, and brand-specific context produces something qualitatively different. Whether that holds up in production — across different industries, brand sensibilities, and edge cases — is the question no launch-day case study can answer.
The founders have exited multiple companies. Theyve shipped to real scale before. That history is the best evidence the product will mature beyond the demo.
