How to use Codex to test flight without pausing time

Adam Blumoff built a flight simulation with Codex, then evaluated fresh agents through WebMCP. An early version gave the agent too little control: hidden simulator logic handled much of the flight. A later attempt appeared successful partly because the agent stopped time while thinking, undermining the evaluation.

Blumoff pushed the design toward a continuous world. The simulator advances at a fixed rate outside React state. The agent reads the aircraft manual and writes commands with activation conditions before takeoff, then replaces its program after a sealed emergency. Simulated air traffic control supplies routes, while the agent chooses how to fly them.

Ten cold-thread attempts initially failed. A manual clarified aircraft limits and command contracts without revealing the emergency or providing a flight program. The resulting version completed three consecutive fresh blind runs without pausing or changing the simulation rate. Trajectory exports retain tool calls, state changes, timing and outcomes.

The human checkpoint was evaluating whether the agent actually owned decisions and whether time kept moving. This is a research prototype with a simplified aircraft model, not a certified training device or evidence that an agent can control a real airplane.

Prompt
Original prompt

Source: Devpost · Adam Blumoff