LangChain released a Slack design agent that costs 16 times less

LangChain released an open-source design agent on 27 October. A marketing or design team can ask for a slide deck, a chart, a social graphic or a short video. The request goes in a Slack thread. The agent sends the finished file back to the same thread. The code is in a public GitHub repository called langchain-ai/design-agent. It is MIT-licensed in practice through Apache 2.0.

The news is in two parts. The team shared how they built the agent in a long blog post. They also open-sourced the whole system. Amal Irgashev and Shreya Talegaonkar wrote the post.

What the team says about use so far

The post lists numbers from the team's own use inside LangChain. The agent has made more than 3,000 assets. It has produced more than 80 slide decks. Posts carrying its work had 4.5 million views on X. Those posts had about twice the likes and 3.7 times the bookmarks of posts without the agent.

The post also tracks user behaviour inside LangChain. 75 percent of users come back. 75 percent of requests now come from staff outside marketing.

The cost per request fell from about $8 in mid-September to about $0.50 now. The post does not break the gain into the parts that moved it. LangChain says the drop came from caching, from reusing finished skills, and from doing less work with the largest model.

How it is put together

The agent runs on LangChain's own stack. Deep Agents is the loop and the sub-agent planner. Managed Deep Agents is the paid service that runs it. LangSmith Sandbox gives each Slack thread its own clean Linux box. The agent writes code in that box to make each file.

The team split the agent into 48 skills across about 1,000 files. One router skill picks the skill to run. A memory system keeps a shared team layer and a private user layer. Both are capped at 200 lines, which keeps the main model from reading context that no longer matches the task.

Video uses HyperFrames, HeyGen's framework for stitching clips with GSAP timelines. Charts use ECharts, rendered by Chrome Headless Shell. Design canvases use a tool called Paper.

What LangChain says it still gets wrong

Early output looked polished but generic. The team calls this "zombie UI". The agent needed a lot of calibration on brand rules. It also had to learn why a human reviewer had rejected a draft. The post does not say an eval set catches this. The team still watches it by hand.

Source

LangChain, Amal Irgashev and Shreya Talegaonkar. Repository: langchain-ai/design-agent.