Coding agents can finally look at the app they just changed

01

WHAT ACTUALLY CHANGED

Compose Multiplatform 1.12 includes an experimental MCP server inside Compose Hot Reload. A compatible coding agent can trigger a reload, capture a screenshot, inspect the interface’s semantic tree, simulate clicks and text entry, and read application logs.

That creates a feedback loop that resembles a very patient tester. The agent changes a button, reloads the app, looks at the result, checks whether the button exists in the semantic structure, clicks it, and reads the logs if something breaks. The developer can review the same visible evidence instead of trusting a confident summary.

The semantic tree is especially useful because pixels alone are ambiguous. It describes meaningful interface elements such as labels, buttons, fields, and states. Screenshots show what a human sees; the tree gives the agent a structured map of what the interface believes it is.

02

WHY THIS MATTERS

Coding agents have been strongest where the result can be checked by a compiler or test suite. Interface work is harder because a change can be valid code and still look wrong, hide a control, or fail after one click. Visual and interactive inspection gives the agent another source of truth.

This does not remove the need for human taste or accessibility review. It does make routine feedback cheaper. The useful shift is from an agent that edits files to an agent that can observe the consequence of those edits and leave a small receipt of what it tested.

03

WHERE IT COULD HELP

• Ask an agent to repair a layout and verify the rendered result
• Test a short click path after changing navigation or form logic
• Inspect labels and controls through the semantic tree
• Capture screenshots and logs as evidence for a human code review