How LangChain Built an Open-Source Design Agent in Slack

LangChain's account of building Content Studio, an internal design agent that takes requests in Slack and returns finished videos, graphics and slide decks, credits the agent's biggest improvement to capturing the company's design judgment in an environment the agent could learn from, alongside better underlying models. The model could already write HTML, chart specs and animation timelines, LangChain says, but without the company's context it reached for familiar patterns, so the work came out polished and generic. The fix was layered. Brand basics such as color tokens and spacing came first, then design principles like one focal point per frame, guardrails against common AI design patterns, and the reasoning behind designers' feedback, captured in skills the agent reads on later jobs. Anything measurable, such as wrong default capitalization, moved into deterministic checks written in code, which LangChain calls gates.

The rest of the post describes the agent's working environment. Each Slack thread gets its own LangSmith Sandbox, an isolated Linux machine where the project persists for the life of the thread. Every asset is built as code, a slide as an HTML page or a video as an animation timeline, so a request to bring a character in earlier becomes a change to one timing value and a re-render rather than a regenerated asset. Long jobs are split across subagents that run in the background while the main agent coordinates them and keeps the thread updated with a plan, live progress and an approval card before anything is published. LangChain cites an overnight reskin of a 52-slide deck in which the main agent ran for about an hour while its subagents together worked more than six hours, without blocking the conversation.

The lesson most likely to carry over to other agent builders is about which instructions age. LangChain's first version tried to do the designing for the model, with 115 reusable components for titles, layouts and animations. That kept work on brand but repeated the same layouts frame after frame. When Opus 5.5 arrived, the team says quality jumped, and removing most of the hard-coded rules improved results further, with more variety. LangChain's conclusion is that rules written to compensate for a model's weaknesses expire, while design principles, tools and context keep their value.

The results are LangChain's own figures about internal use: more than 3,000 assets delivered, three in four requests now coming from outside marketing, and a median request cost of about 50 cents in model spend, down from about $8 in mid-September. The post does not attribute that cost drop to any single change and reports no outside evaluation. LangChain says the full harness is open source and can be deployed to a team's Slack through its Managed Deep Agents service.