LangChain Put a Design Agent in Slack; the Hardest Work Was Deleting Rules

LangChain recently open-sourced Content Studio, a design agent that lives in Slack. You drop in a request, and it comes back with video, social graphics, slides, or even redrawn UI mockups. Every asset is written as code.

LangChain shared a short clip it made itself: a parrot standing on a branch in a pixel-art jungle. The easiest thing to miss in the video is that the parrot can move to the left. It doesn’t have to regenerate the whole asset; you change one time value and render again.

The thinking behind this is worth unpacking. Many design-agent demos stop at generating an image. LangChain spent four months on something else: making the output look like its own company’s work and making it continuously editable.

Content Studio has already run more than 3,000 assets and 80-plus slide decks internally. Posts on X featuring its visuals got about 4.5 million views. On the LangChain account, these posts got about 2x the likes and 3.7x the bookmarks. Three-quarters of requests came from outside marketing, and 75% of users come back every week. The median request costs about 50 cents in model spend; in mid-September it was still about $8. After migrating to Managed Deep Agents, their own code shrank by about 80%.

Those numbers look good, but the more interesting part of the article is the process. Early feedback was bad. LangChain’s GTM engineering team and design team worked together to write designers’ judgment into the agent’s skills and checks. Anthropic calls the model’s tendency to default toward common patterns distributional convergence. Stripe design lead Katie Dill warned of a similar risk with zombie UI: quickly copying modernism’s clean surface while losing the thinking behind it. Her warning is that AI will do the same thing—you have to have your own point of view. LangChain puts it more directly: the model already knows how to write HTML, CSS, chart specs, and animation timelines, but it doesn’t have your context, so on every decision the brief doesn’t cover, it fills in the answer it has seen most often.

They gave Content Studio more than brand colors and fonts. They also gave it design principles, such as one focal point per frame; guardrails against AI clichés; and the reasoning behind every piece of feedback. One example was early motion design: a bird flies out of frame, trailing a light streak, with the requirement no sharp edges. The first version had hard edges. The designer meant that the edges should be soft, the tail should taper and fade out, and the direction should run opposite the bird’s movement. Later, the agent would sketch on its own to locate the bird’s tail, then do frame-by-frame checks to make sure the base of the light streak stuck to the tail. Most viewers won’t notice this detail, but if it’s wrong, it feels off.

Another counterintuitive point: rules expire. They first built 115 reusable UI components, hoping to help the model make fewer creative decisions. As a result, the agent repeated the same layout and animation every frame. After Opus 5.5 came out, quality jumped a level, and they went back and deleted most of the hard-coded rules, keeping only design principles and high-level guidance—and the variation increased. LangChain calls this its own Bitter Lesson. Rules that patch the model’s weaknesses become obsolete as the model gets stronger; tools, context, and how people guide the work are more durable.

In the comments, John Rood raised a more practical question: you can delete rules, but every rule needs to carry its why. Many rules are scars left by one bad run; a month later they look like arbitrary requirements, and if you delete them, you’ll step on the same rake again. Huilong also warned that one-off exceptions like use neon just for this launch shouldn’t be absorbed into shared rules by the slow learning loop. These two issues show that memory and rules aren’t better when there are more of them; you have to distinguish company standards, personal preferences, and one-off exceptions.

The harness part. Each Slack thread gets its own LangSmith Sandbox, an isolated Linux machine with a filesystem and shell, and the project persists for the lifetime of the thread. Slides are HTML, charts are data specs, and videos are animation timelines. Moving the parrot left means changing one value. On the tool layer: HyperFrames is HeyGen’s open-source video framework that turns video into a web page and uses GSAP timelines to move elements; motion-studio is an in-house canvas engine that draws pixels frame by frame, mixing up to 16 moments per frame to create shutter blur. A storyboard frame takes about 50 ms, while a second of final footage takes 8 to 40 seconds, so the agent first tries directions cheaply in the storyboard, then makes an animatic after approval, and then goes into production. The Paper canvas lets the agent control files directly through MCP, and the team can compare directions side by side in the same file and leave comments on specific frames. To run Paper Desktop in a cloud sandbox, they run it headlessly on an always-on Linux worker, using a virtual display and auth bridging.

Long tasks rely on asynchronous sub-agents. The main agent is more like a chief of staff: it breaks down tasks, assigns work, and runs things in parallel. In Slack it shows the plan, ETA, run status, render progress, cost, and approval cards. Tools include start_task, check_tasks, cancel_task, resume_task, report_progress, and schedule_wakeup. One overnight task was redoing a 52-page slide deck; the main agent ran for about an hour, the sub-agents worked more than six hours combined, and the conversation wasn’t blocked.

Memory has two loops. The fast loop saves rules in the same turn the user gives feedback, split into team-shared and private layers; each layer has an AGENTS.md that is loaded on every model call. Early files grew to 500,000 tokens; now they’re limited to a 200-line index. The slow loop uses learn_from_job to spin up a lightweight subagent after a task, records at most five evidence-backed lessons, and links them to the skills used; future tasks read and rate them, and a weekly background agent merges the useful ones into skills and the wiki.

Now the repo is open source, including the sandbox computer, skills, wiki, memory, asynchronous sub-agents, and Slack UX. You can fork it, add brand tokens, fonts, and skills, and deploy it to your team’s Slack with Managed Deep Agents. For designers, the role looks more like creative director: set direction and review work. For PMs, engineers, and founders, they can drop requests to the agent in Slack and take away production-ready assets.

Many agent products are still proving they can generate. The judgment in LangChain’s retrospective is calmer: generation is only the starting point; being editable, reviewable, and able to remember why is what makes a company willing to use it every day.