AI Trends

LangChain Open-Sources Design Agent, Bringing Its Internal Content Studio to Slack Teams

Shao MengUpdated Nov 20 10:59

English text is being prepared.

AI summary

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

Tags

Quoted post

Harrison Chase

@hwchase17

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

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

View original post

Body

LangChain 开源 Design Agent,把公司在用的 Content Studio 公开,用每线程沙箱和品牌检查在 Slack 交出平面、演示文稿和影片

@LangChain 把公司内部在用的 Design Agent「Content Studio」开源。它建立在 Deep Agents 上,默认用 Managed Deep Agents 部署到团队 Slack:在线程里说明需求、附上文件,就会拿回符合品牌规范的平面、多尺寸广告、演示文稿、图表、分镜和影片。也可以用命令行、HTTP API,或自己用 Docker 托管。自 8 月起它已为公司交出 3000 多件物料,演示文稿超过 80 份;带这些视觉的帖文在 X 上大约有 450 万次浏览。四分之三的需求来自市场部门以外。单次请求的模型费用中位数大约是 0.5 美元,9 月中旬还大约是 8 美元。

开源地址:https://t.co/wDwS1cTTF3

一条 Slack 线程、一台设计用电脑

Managed Deep Agents 负责 Slack、记忆、定时任务,以及每条线程一台 LangSmith 沙箱;仓库补上设计工作流、品牌包和渲染工具。沙箱是一台 Linux 机器,里面有 Chrome、FFmpeg、HeyGen 的 HyperFrames、D2 和团队字体,由 sandbox/setup.sh 搭起来。成品都以代码写在项目目录里:幻灯片是 HTML,影片是动画时间线,改完再渲染即可。密钥不进沙箱,默认也没有外网;图像、语音、音乐在沙箱外调用,Notion、GitHub、Figma 一类 MCP 经 MDA 接入,每个人用自己的账号登录。

一条 Slack 消息会随步骤更新,页脚显示进度和目前费用,中途补的意见也能进正在进行的工作。大任务由主 Agent 拆开,在同一沙箱里后台启动子 Agent;做完、失败或卡住时,MDA 会叫醒主 Agent,它不用空等,还可以继续对话。影片按分镜、动态预演、正式制作推进,每一段先过检查再组装。HyperFrames 把影片当成时间线上的网页来做,Paper 是可以改分镜或版式的共享画布。

技能路由、品牌检查,以及它怎么变好

常驻提示很短。每次请求先读路由技能,再从 35 项技能里选出对应流程,由 Context Hub 按需加载。设计维基里是字体、色彩、版式、图表这类手艺,品牌包里是颜色、字体、标志和语气。渲染之后会先看成品再发出;颜色、字体、对比、尺寸和图表计算由代码在每次渲染时检查。规则可以说一次就记下,分成全员共享的团队记忆和每人一份的个人记忆。任务结束后,learn_from_job 按所用技能记下有依据的经验,后续任务会阅读并评分,站不住的会被丢掉。

他们写到,早期为了补模型短板做过 115 个可复用界面组件;Claude Opus 5.5 出来后质量明显提高,他们删掉了大部分硬编码规则,留下设计原则和高层指引。默认模型就是 Claude Opus 5.5,也支持 OpenAI、Google 和 LangSmith Gateway。Managed Deep Agents 在 LangSmith Cloud 上仍是公开测试;代码免费,模型和托管要另付。本地可以用脚本化模型、不配密钥跑演示,走真实的 Agent 图、沙箱和检查。

Source

Shao Meng