LangChain's design agent trades hard-coded rules for context: median request cost falls from about $8 to about 50 cents

The quality of a design agent depends mainly on the context it is given, not on whether the model can write code; the model already handled HTML, CSS, chart specs and animation timelines but lacked the company's judgment and fell back on what it had seen most. Early feedback was not great, and output could look polished while resembling a hundred others, or be beautifully animated without making its point. On the same prompt, a local coding agent with the brand skill still missed the correct logo lockup, text alignment and primary versus secondary brand colors, while Content Studio produced a fully on-brand image.

The agent works on a designer's computer: each Slack thread gets its own isolated Linux sandbox with a filesystem and a shell, and the project lives there for the life of the thread. Long, iterative design work needs a place to keep state across many turns rather than restarting each time. Everything is written in code, including HTML for slides, data specs for charts and animation timelines for video, so changing a single timing value and re-rendering leaves artwork and copy untouched.

Design knowledge is split into four parts: skills that say how to do a job, a wiki that supplies company and brand context, deterministic checks written in code, and a memory layer that records what people teach the agent. The first harness tried to do the designing for the model, including 115 reusable UI components, and the agent repeated the same layouts and animations for each frame. After Opus 5.5 arrived, the team removed most hard-coded rules and kept design principles and high-level guidance, and results improved further with more variety and less repetition.

Long jobs run on asynchronous subagents while the main agent acts as chief of staff, showing the plan, live progress and an approval card in the Slack thread before anything ships. Synchronous subagents force the main agent to wait, leaving no way for people to interact while a film is being built. In an overnight job to reskin a 52-slide presentation, the main agent ran for about an hour while its subagents collectively worked more than six hours without blocking the conversation.

The setup runs inside the company's Slack environment and covers code-written assets such as video, social graphics, recreated UI images and slide decks; the harness is open-sourced so others can add their own brand tokens, fonts and skills and deploy it. The team says the next step is for the agent to manage the entire campaign lifecycle, writing the brief, making assets per channel, scheduling posts and using performance data for the next campaign, with people approving everything before it goes public.

The team reports that early feedback was not great and that rules compensating for model weaknesses expire as models improve, which is why most hard-coded rules were removed; subjective calls such as whether a campaign idea is strong enough still come from people and are captured in skills. The cost and engagement figures are internal company statistics, with no external control or cross-team replication reported.