LangChain Open-Sources Design Agent, Bringing Its Internal Content Studio to Slack Teams
LangChain has open-sourced Content Studio, its internally used Design Agent, which can generate brand-compliant graphics, multi-size ads, presentations, charts, storyboards, and videos within Slack threads. It also supports a CLI, HTTP API, and Docker self-hosting. Built on Deep Agents and Managed Deep Agents, the system assigns a LangSmith sandbox to each thread and organizes workflows through 35 skills, brand packs, and rendering checks. Since August, it has delivered more than 3,000 assets, while the median model cost per request has fallen from about $8 in mid-September to around $0.50.
Why read this
SelectedAI score 84
The materials showcase the design agent’s complete workflow, from task decomposition and sandbox isolation to brand compliance checks, offering teams a reference for deploying similar systems.
Topics
View original postfun fact: as we were gearing up towards launch, we had our content agent take a stab at making a launch video we didn't end up using it - but it was really good and we definitely could have! attaching below this agent is Open Source as well! https://t.co/t5XHzTow2O https://t.co/GQpKp6MlyY
View original
fun fact: as we were gearing up towards launch, we had our content agent take a stab at making a launch video we didn't end up using it - but it was really good and we definitely could have! attaching below this agent is Open Source as well! https://t.co/t5XHzTow2O https://t.co/GQpKp6MlyY
Body · AI translation
LangChain open-sources Design Agent, making its in-house Content Studio public to deliver graphics, presentations, and videos in Slack with per-thread sandboxes and brand checks
@LangChain has open-sourced Content Studio, the Design Agent used internally by the company. Built on Deep Agents, it is deployed by default to team Slack workspaces using Managed Deep Agents: describe what you need in a thread and attach files, and it will return brand-compliant graphics, multi-size ads, presentations, charts, storyboards, and videos. It can also be used through the command line or an HTTP API, or self-hosted with Docker. Since August, it has delivered more than 3,000 assets for the company, including over 80 presentations; posts featuring these visuals have received roughly 4.5 million views on X. Three-quarters of requests come from outside the marketing department. The median model cost per request is about $0.50, down from roughly $8 in mid-September.
Open-source repository: https://t.co/wDwS1cTTF3
One Slack thread, one computer for design
Managed Deep Agents handles Slack, memory, scheduled tasks, and one LangSmith sandbox per thread; the repository adds design workflows, brand kits, and rendering tools. Each sandbox is a Linux machine equipped with Chrome, FFmpeg, HeyGen’s HyperFrames, D2, and the team’s fonts, set up through sandbox/setup.sh. All deliverables are written as code in the project directory: slides are HTML, while videos are animated timelines, so they can simply be rendered again after edits. Secrets never enter the sandbox, which also has no external network access by default; image, voice, and music services are called outside the sandbox, while MCPs such as Notion, GitHub, and Figma are connected through MDA, with each person signing in using their own account.
A Slack message updates as the work progresses, with the footer showing progress and the current cost; feedback added midway can also be incorporated into work already in progress. For large tasks, the primary Agent breaks the work down and launches sub-Agents in the background within the same sandbox. When they finish, fail, or get stuck, MDA wakes the primary Agent, so it does not have to wait idly and can continue the conversation. Video production proceeds through storyboarding, animatics, and final production, with each segment checked before assembly. HyperFrames treats a video as a web page laid out on a timeline, while Paper is a shared canvas for editing storyboards or layouts.
Skill routing, brand checks, and how it improves
The persistent prompt is short. Each request first reads the routing skill, then selects the appropriate workflow from 35 skills, which Context Hub loads on demand. The design wiki covers craft such as typography, color, layout, and charts, while the brand kit contains colors, fonts, logos, and tone of voice. After rendering, the system inspects the finished work before sending it; code checks colors, fonts, contrast, dimensions, and chart calculations on every render. A rule only needs to be stated once to be remembered, either in team memory shared by everyone or in a separate personal memory for each individual. After a task ends, learn_from_job records evidence-based lessons under the skills used. Subsequent tasks read and score those lessons, discarding any that do not hold up.
The team wrote that it initially created 115 reusable interface components to compensate for the model’s shortcomings. After Claude Opus 5.5 was released, quality improved substantially, and they removed most of the hard-coded rules, retaining design principles and high-level guidance. Claude Opus 5.5 is the default model, though OpenAI, Google, and LangSmith Gateway are also supported. Managed Deep Agents remains in public beta on LangSmith Cloud; the code is free, while model usage and hosting cost extra. Locally, a demo can run with a scripted model and no configured secrets while still using the real Agent graph, sandbox, and checks.
AI Trends