LangChain open-sources Slack design Agent Content Studio and its harness

Putting design taste into an Agent: LangChain's four-month practice with the open-source Content Studio

LangChain's internal GTM engineering and design teams spent four months taking Content Studio (the Design Agent in Slack) from "ugly, like cookie-cutter AI drafts" to production-ready: writing designers' judgment into skills / wiki / code gates, giving the model a sandbox computer that can revise projects across turns, and then using asynchronous sub-Agents + a collaborative UX with multiple people on a thread to sustain multi-hour tasks; they also open-sourced the entire harness, which can be forked and hooked up to your own brand, then deployed to your team's Slack via Managed Deep Agents.

Open-source address: https://t.co/wDwS1cTTF3

Project background and results The company moves fast, and launches, events, blogs all need images, clips, intros, and ad materials; design headcount was fully occupied by repetitive requests. Content Studio: describe the request in Slack, attach a brief/assets, and get back a compliant finished piece. The article says that since August it has delivered 3,000+ assets (including 80+ PPTs); posts with its assets on X got about 4.5 million impressions; on the official account, such posts got about 2× the likes and 3.7× the saves of other posts; about 3/4 of requests came from teams outside marketing, and 75% of users returned weekly. Median model cost per request dropped from about $8 in mid-September to about $0.50; after migrating to Managed Deep Agents, the amount of their own code was reduced by about 80%.

Can AI design? Models can already write HTML/CSS, chart specs, and animation timelines; what they lack is a stance and context, and they easily "converge in distribution" into good-looking but faceless drafts (the article cites Anthropic's claim, as well as Stripe design lead Katie Dill's "zombie UI": without a point of view, AI will supply one for you). The conclusion is not "AI can't design," but that you must teach it your taste.

How to teach taste? · Basics: color palette, font weights, spacing, alignment. · More critical: design principles (such as one focal point per frame), guardrails against AI clichés, and the reasoning behind each piece of feedback (not just "change it"). · Measurable things use code gates (such as default capitalization); subjective judgments (whether the creative is strong enough, whether the information is clear) are decided by designers and then written into a skill for later reuse. · Early motion example: a bird flying out trailing a light ribbon, where "no sharp corners" was understood as hard-cutting rounded corners; later the Agent would build its own frame-by-frame checks to ensure the light ribbon hugs the motion. The designer's role is more like a creative director: set direction, review drafts, rather than hand-crafting every piece.

Media harness: the designer's computer The foundation is Deep Agents (planning, files, sub-Agents, context management); Managed Deep Agents handles production operation (Slack, LangSmith Sandbox, tools, memory, scheduling). Environment-side priorities: · One sandbox per Slack thread: an independent Linux + file system, with the project persisting across turns; output is code as much as possible (slide HTML, animation timelines), so changing a time point re-renders without regenerating the whole thing. · Tool layer: editing uses HeyGen's open-source HyperFrames (treating the film as a web page + GSAP timeline); original animation uses the in-house motion-studio (drawing frames pixel by pixel, mixing multiple moments to simulate shutter blur, etc.); audio and video share the same timeline. · Four layers of knowledge: Skills (routed by task; the article says about 48 skills and nearly a thousand files; the README says about 35, subject to the repository's current state); Wiki (brand, tone, products, breakdowns of about 70 reference films, etc.; skills only link to it rather than copy it); Gate (deterministic checks); Memory (team layer + personal layer; AGENTS.md's index once ballooned to about 500k tokens, a lesson learned, and is now limited to about 200 lines). · Spec-driven: the expensive part is the finished piece, so first a storyboard (exploration on the order of about 50ms/frame) → shot breakdown/animatic to check pacing → then mass production; the order is enforced by code, while judgment stays in the prompt. The collaborative canvas uses Paper (click to change, rather than describing coordinates in natural language the whole time); Paper's desktop capability runs headless on a persistent Linux worker and is bridged to the cloud sandbox.

The main Agent coordinates and assigns work On long tasks, multiple people revise the brief as they go. Synchronous sub-Agents would block the main thread, so subtasks are asynchronous: the main Agent breaks down work, assigns roles, chooses models and effort levels, and coordinates with start_task / check_tasks / cancel_task / resume_task / report_progress / schedule_wakeup; human messages can be inserted before the next round of model calls. The thread continuously updates the plan, ETA, subtask status, and cost, and approval is required before publishing. Example: overnight retouching of a 52-page PPT, where the main Agent ran for about 1 hour and sub-Agents totaled more than 6 hours, without blocking the conversation.

Bitter Lesson Early on, they piled about 115 reusable UI/motion components and detailed rules onto weaker models; the brand became stable but layouts repeated. After the models (the article says Opus 5.5) jumped, they deleted most hard-coded patches, keeping only principles and high-level guidance, and diversity actually increased. Rules that compensate for model shortcomings expire; what lasts longer is the environment: tools, context, and how humans intervene midway.

Self-improvement Fast loop: write rules into memory on the spot. Slow loop: learn_from_job extracts at most 5 evidence-backed lessons and attaches them to the corresponding skill, later tasks are scored, and every week a background process merges them into skills/wiki. Next direction: the entire campaign lifecycle (posting via Buffer after approval, employees sharing drafts, scheduled reviews, etc.), with humans still reviewing and approving.

Open source and deployment design-agent: after forking, add brand tokens, fonts, and skills, and deploy to Slack with Managed Deep Agents. The README emphasizes that it can be tried locally and that keys do not enter the sandbox; the architecture diagram strings together Slack → main Agent → skills/wiki/memory → sandbox (Chromium, ffmpeg, HyperFrames, motion-studio, asynchronous sub-Agents).