How the agent tried to land
29 practice attempts, then 8 final flights. None landed safely.
The setup
The agent can only see the cockpit video. It writes a controller, flies for 20 seconds, watches what happened, and edits the code. Open any attempt below to watch along.
MODEL USE 8,873,990 tokens$4.58
Cached: 8,754,648
Prompt: 87,146
Output: 32,196
No trajectory steps match that filter.