LangChain's Slack Design Agent Shipped 3,000 Assets as the Typical Request Fell from $8 to 50 Cents
LangChain did not have enough designers for all its requests, so the team built Content Studio, an agent that lives in Slack. Given a description, it produces videos, slides, and graphics in the company's brand style. It has shipped more than 3,000 assets, and the typical (median) request now costs about 50 cents in model fees, down from $8 in mid-September. The takeaway: keep your design principles, and delete the crutch rules when a stronger model arrives.
AI-processed from LangChain Blog; edited by Hamidun News
Every startup team runs short of designers. Launches, webinars, blogs, and events all need headlines, charts, presentations, and videos, and designers drown in repetitive work. LangChain tackled this problem with Content Studio, an agent that lives in Slack. A person describes the task and attaches a brief, and the agent returns finished files in the company's style. Over four months it has shipped more than 3,000 assets, and the typical (median) request now costs about 50 cents in model fees. In mid-September, the same request cost about $8. The team explained how the system works and opened up the code.
What happened in four months
Among the 3,000 assets, more than 80 are presentations. One of them became the company's most-watched organic video on YouTube. Posts about the agent's work on X collected about 4.5 million views, and LangChain's posts featuring its graphics get roughly twice as many likes and 3.7 times as many bookmarks as the rest.
Three out of four requests now come from outside marketing: product managers, engineers, and founders use the agent. 75% of users return every week. After the move to Managed Deep Agents, the team's own code shrank by roughly 80%. The service now handles the managed runtime, the Slack integration, the agent's memory, and scheduled runs.
The start was rough. According to the authors, the first reviews were "not great," and it took four months to teach the agent taste, mostly through mistakes made in full view of the whole company.
How to teach an agent taste
Models were already good at writing HTML, CSS, chart descriptions, and animation. But without context, they chose the most familiar solutions. The result looked neat but resembled hundreds of other works. The authors cite Anthropic's term "distributional convergence" and the "zombie interface" idea from Stripe's head of design: the form is copied, while the thinking and character are lost.
The article gives an example. A local coding agent was given a brand book with colors, fonts, and logos, but it mixed up the logo variant, the text alignment, and which colors were primary and which were secondary. The same request in Content Studio produced an image with no errors. When a brief leaves something unspecified, a generic agent falls back on the most common pattern it has seen, while Content Studio falls back on the company's own decisions.
The rules cover basic things: colors, font weights, spacing, alignment. Then the team added principles such as "one focus per frame," protection against typical AI design tics, and an explanation of why a designer gave a particular piece of feedback. Anything measurable is checked by automation. For example, a script now catches incorrect default capitalization. Subjective qualities, such as the strength of an idea or the clarity of a message, are judged by designers, and their reasoning is written down as skills.
The story of the bird in an animation is telling. The agent was asked for "no sharp edges," but it produced rough edges. A designer would know that the light trail behind the tail should soften and fade in the direction opposite to the motion. Later, the agent checked this detail itself: it sketched a draft and wrote a frame-by-frame check confirming that the root of the ribbon stays attached to the bird's tail until the bird leaves the frame.
A designer's computer and removing extra rules
A designer works in several tools at once: graphics software, a video engine, a shared canvas, plus the company's brand book and documents. The agent needs the same set. Each Slack thread gets a separate isolated Linux machine with a file system and a shell, and the project lives there for the whole conversation. The agent can pick up work between messages, use the video engine, graphics tools, and the shared canvas, and then review its own renders.
For large tasks, asynchronous subagents work, sometimes for hours. That is why the thread shows the plan, the progress, and a request for confirmation before publishing, and any colleague can step in and redirect the work while it runs.
The main lesson concerns rules going stale. At first the team tried to do the design for the model: it built a system of 115 ready-made components for titles, layouts, and animations. This kept the brand consistent, but the agent kept repeating the same layouts. Then Opus 5.5 came out, and quality rose sharply. The team removed most of the hard rules, kept the principles and general guidance, and variety increased. Rules that compensate for a model's weaknesses lose their value over time. What remains are the tools, the context, and the way people steer the work.
A few caveats: this is one team's experience with its own brand, and the numbers are published by the company itself. But the approach transfers to other products. Give the agent a working environment, write down principles and the reasons behind corrections, automate what can be measured, and with each new model check which rules are still needed. The brand colors, fonts, and skills are added to the open repository, and the agent can be deployed to Slack through Managed Deep Agents.
- Nov 21, 2026Anthropic mathematician says Claude helped prove the Riemann hypothesis in six weeks
- Nov 21, 2026OpenAI Opens ChatGPT Spaces to Plus Subscribers: Shared Documents with AI
- Nov 21, 2026Chinese Tech Giants Double AI Spending to $139B, While Labs Recover Less Than 8%
- Nov 21, 2026Ego lite Gives Each AI Agent Its Own Workspace Space in the Browser
Key Questions Answered
What happened in four months?
Among the 3,000 assets, more than 80 are presentations. One of them became the company's most-watched organic video on YouTube. Posts about the agent's work on X collected about 4.5 million views, and LangChain's posts featuring its graphics get roughly twice as many likes and 3.7 times as many bookmarks as the rest. Three out of four requests now come from outside marketing: product managers, engineers, and founders use the agent. 75% of users return every week.
How to teach an agent taste?
Models were already good at writing HTML, CSS, chart descriptions, and animation. But without context, they chose the most familiar solutions. The result looked neat but resembled hundreds of other works. The authors cite Anthropic's term "distributional convergence" and the "zombie interface" idea from Stripe's head of design: the form is copied, while the thinking and character are lost. The article gives an example.
What are the key details on A designer's computer and removing extra rules?
A designer works in several tools at once: graphics software, a video engine, a shared canvas, plus the company's brand book and documents. The agent needs the same set. Each Slack thread gets a separate isolated Linux machine with a file system and a shell, and the project lives there for the whole conversation. The agent can pick up work between messages, use the video engine, graphics tools, and the shared canvas, and then review its own renders.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.