An AI agent just flew a US military jet for 17 hours

Top Gun pilots might want to look away. Military flight controls are officially being handed over to neural networks.

During a test campaign in December, an AI agent piloted Lockheed Martin’s VISTA X-62A tactical training aircraft for over 17 hours. According to Lockheed, this marks the first time an artificial intelligence system has flown a tactical jet in this manner.

Skunk Works builds a physical simulator

VISTA stands for Variable In-flight Simulation Test Aircraft. Its original mission was using software to mimic the handling and flight characteristics of other jets. This time, however, the system was configured to mimic a human pilot in real skies.

Developed by Lockheed Martin’s Skunk Works in tandem with Calspan, the aircraft lives at Edwards Air Force Base with the US Air Force Test Pilot School. It is not an unproven sandbox. The platform was previously used to mature the virtual heads-up display and voice command systems for the F-35 Lightning II.

The Air Force recently upgraded the X-62A with Lockheed’s Model Following Algorithm and the System for Autonomous Control of the Simulation. These upgrades essentially turn the jet into a flying testbed for autonomy, allowing teams to load up experimental code and validate edge-case maneuvers in real physics.

Decoupling airframes from algorithms

Dr. M. Christopher Cotting, director of research at the USAF Test Pilot School, explained that VISTA allows the military to parallelize software and hardware. Instead of waiting for a bespoke uncrewed jet to roll off the assembly line before writing flight software, engineers can develop cutting-edge AI on VISTA right now.

By the time new uncrewed combat air vehicles are manufactured, the autonomy stack will already have logged actual airtime. The goal is to compress development timelines and push battle-ready autonomous features straight to warfighters.

Why it matters

Logging 17 hours on a tactical aircraft is a massive proof point for the software-defined warfare playbook.

We think Lockheed Martin’s architecture here is particularly shrewd. Designing brand-new unmanned tactical jets from scratch is outrageously expensive, and testing green autonomous models on unproven airframes carries high crash risks. By using an existing, highly adaptable platform that can simulate multiple aerodynamic profiles, the Air Force gets real-world aerodynamic noise and sensor reality checks without betting the farm on experimental airframes.

For defense contractors, the battle lines are shifting from purely mechanical engineering to rapid algorithmic training loops. The winners will be teams that decouple flight software from physical metal and iterate autonomy stacks like modern software companies.

That said, 17 hours of controlled flight is still miles away from dogfighting. Keeping an aircraft within safe envelopes is one thing; navigating electronic warfare, sensor degradation, and split-second tactical decisions is another. Autonomy has mastered the stick, but combat judgment remains a much harder problem.