AivexaNewsSearch
AI news for builders and product teamsChecked every hour

How We Built Our Design Agent

Collected Oct 27, 2026

LangChain has published a detailed account of building Content Studio, an in-house design agent that lives in Slack and turns a single prompt into videos, social graphics, recreated UI images and slide decks. The GTM Engineering team built it with the design team over roughly four months, running it on Managed Deep Agents, and has now open-sourced the full harness so others can fork the repo, add their own brand tokens, fonts and skills, and deploy it to a team Slack channel. The company will walk through the harness and UX at GTM Engineering Live on October 28.

The motivation was volume. LangChain runs continuous launch campaigns, webinars, blogs and events, each needing headers, diagrams, decks, videos and ad creatives. The design team was the bottleneck, with repetitive requests crowding out strategic work. Content Studio lets anyone at the company describe what they want, attach a brief or content, and receive finished on-brand files. Every asset the agent produces is written in code rather than generated as a flat image or clip.

The reported results are concrete. Content Studio has delivered more than 3,000 assets, including over 80 slide decks, one of which appeared in LangChain's most-watched organic YouTube video. Posts featuring its work drew about 4.5 million views on X, and on the LangChain account posts with its visuals earned roughly 2x the likes and 3.7x the bookmarks of other posts. Three in four requests now come from outside marketing, and 75% of users return weekly. The median request costs about 50 cents in model spend, down from about $8 in mid-September, and LangChain's own code shrank by about 80% after migrating to Managed Deep Agents.

LangChain's central argument is that models already execute well — writing HTML and CSS, chart specs and animation timelines — but lack a point of view, so they fall back on familiar patterns. The post cites Stripe Head of Design Katie Dill's comparison to postwar copies of modernism that kept the look while losing the thinking, which she calls "zombie UI." The fix was teaching taste: starting with color tokens, font weights, spacing and alignment, then adding design principles such as one focal point per frame, guardrails against common AI design patterns, and the reasoning behind each piece of feedback. When a designer makes a subjective call, the reasoning is captured in a skill so the agent can apply it later.

An early motion graphic illustrated the gap. Asked for "no sharp edges," the first pass produced harsh, rounded edges; a designer would have understood the ribbon should soften while its tail tapers and fades opposite the bird's movement. In a later iteration the agent sketched where the tail should be and built its own frame-by-frame check to confirm the ribbon stayed attached. Measurable problems were easier: default capitalization is now caught by an automated check.

The harness gives the agent a designer's computer. Each Slack thread gets its own LangSmith Sandbox, an isolated Linux machine with a filesystem and shell where the project persists for the life of the thread. That matters because design is iterative: if someone asks to bring an element in earlier, the agent can change a single timing value and re-render without touching artwork or copy, whereas a flat image or generated clip would require recreating the whole asset. The brief, plan, scene code, audio and renders all sit in one project folder shared by every subagent on the job.

Motion runs on two engines. HyperFrames, HeyGen's open-source video framework, turns a video into a web page where clips, screen recordings, captions and text are elements moved by a GSAP timeline, so the agent edits a webinar or product demo like a web page; its CLI lints the composition, pulls frames and renders cuts, and HeyGen's official skills teach the format. A second in-house canvas engine, motion-studio, handles what a web page cannot: blur on fast motion, glow, 3D logo turns. The agent writes code that draws every pixel of every frame, and the engine blends up to 16 moments per frame to mimic a camera shutter, which is how the parrot film was made. Both engines render frames by timestamp, so a frame at ten seconds looks the same whenever it is rendered, letting the agent inspect individual moments without rendering the whole film and allowing parallel frame rendering. Audio shares the same timeline, keeping narration, music and effects synchronized.

Knowledge is split into four functions. Skills guide how to do specific tasks: a router skill maps each kind of ask (campaign, restyle, thumbnail, post, YouTube package, launch film, short animation) to a path and specialist skill and decides whether a shared canvas is needed. LangChain has 48 skills spanning close to 1,000 files, each scoped to a job rather than a topic, so descriptions read like routing rules, including when not to use them. Skills compose — launch-film directs while motion-studio makes picture and sound — and launch-film carries role briefs for subagents plus a case study of an 88-second launch film rebuilt shot for shot in code, which taught the agent to split work at hard scene changes, keep continuous camera moves with one subagent, and freeze shared code before parallel work. MCP tools sit behind a keyword search across names and descriptions so extra tool schemas load only when needed. The wiki holds brand tokens, graphic language, voice, platforms, campaigns and product information, plus a motion wiki covering directing, craft, sound and review alongside breakdowns of 70 reference films; skills link to it rather than duplicating it, so a font rule change happens once. A separate memory layer captures what people teach the agent over time.

Cost discipline shaped the workflow. A storyboard frame renders in about 50 milliseconds, versus 8 to 40 seconds for a single second of finished film, so the agent develops ideas in the storyboard where exploration is cheap and moves to full production once a direction is approved. LangChain also describes a Bitter Lesson: its first harness prebuilt a design system of 115 reusable UI components to reduce creative decisions, which kept output on brand but produced repetitive layouts. After the arrival of what it calls Opus 5.5, quality jumped, and the team removed most hard-coded rules while keeping design principles and high-level guidance, gaining variety. Rules that compensate for model weaknesses, it argues, expire; the environment, tools, context and human guidance last.

Collaboration shapes the UX. Big jobs run for hours on async subagents, so the Slack thread shows the plan, live progress and an approval step before anything ships, and anyone on the team can steer mid-run. This suggests that for product teams building agentic tools, the harder work is the sandbox, the context layer and multi-user control surface; model quality is increasingly a given. A likely trade-off is maintenance: skills and wiki content need ongoing curation, and rules written for a weaker model can quietly hold a stronger one back, so periodic pruning may matter as much as authoring. Why it matters: teams that treat a design agent as a container for house style, with versioned context and deterministic checks, can offload repetitive asset production to non-designers — an inference, but one LangChain's stated numbers support — while designers shift toward direction and review.

Read at LangChain

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

How LangChain built Content Studio, an open-source design agent in Slack that makes on-brand videos, graphics and decks, and the harness that powers it.