Mission control for one agent is two URLs

An agent that acts on a GUI has two stories: what it thinks and what it does. Give each one a pipe. The agent tees its reasoning trace to /api/v1/pipe/trace as it plans. ffmpeg captures its display and sends the screen to /api/v1/pipe/screen, watchable with ?video.

Open two tabs. The left one streams the plan: "locating the export button." The right one shows the cursor actually drifting toward the wrong menu. The gap between the two tabs is where your debugging starts.

Observability tools give you traces, and traces are what the model believes happened. The screen is what happened.

One tab shows what it thinks; the other shows what it does.