XNews

Cursor Cloud Agent Rebuilds Slack UI as Code-Generated Video

Cursor Cloud Agent Rebuilds Slack UI as Code-Generated Video

A Cursor cloud agent running Chrome on a Linux remote desktop opened the real Slack and rebuilt its interface as code-generated motion video, without decompiling the app. The agent extracted visual and easing patterns from the live product, produced channel and thread assets, and iterated in a render-and-compare loop. The workflow was codified into a reusable motion skill.

Event version 1.

Key insights

No decompilation involved

The agent was told what to build and given video references, then opened the real Slack on its own computer to extract visual and easing/motion patterns. Slack was not decompiled.

Video as code enables visual verification

Generating video assets as code turns correctness into whether the output renders or compiles and matches expectations. Rendered output is checked by the model against the real reference, then the code is revised.

Inner and outer loops

The inner loop has the agent act, compare against the real thing, fix against clear failure criteria, and continue until it hands off for human review. The outer loop is where humans supply references, preferences and access, and judge artifacts, screenshots and analytics.

The ladder from workflow to abstraction

The path runs from doing a task once, to codifying it as a skill, to generalizing the abstraction beyond Slack to other products, and finally automating the outer loop.

Skills capture failure fixes

When the agent fails, the cause is debugged and written into a skill so later agents inherit the fix and eliminate that class of problem.

Source posts