AI just took the stick of an Air Force combat drone

AI just officially took the stick of an actual combat drone, and it did far more than just keep the wings level.

The U.S. Air Force announced that a stealthy XQ-58A Valkyrie completed a major autonomous flight test at Eglin Air Force Base in Florida on July 25. Powered by an AI/ML agent developed by the Air Force Research Laboratory, the uncrewed jet flew for roughly three hours over the Gulf of Mexico, solving what the service called a tactically relevant challenge problem in mid-air.

Air Force Col. Tucker Hamilton, head of AI Test and Operations, called the mission a critical breakthrough that proved a multi-layer safety framework and officially enabled the development of AI agents capable of modern air-to-air and air-to-surface skills.

Millions of simulated dogfights

Teaching an AI to fly by putting it straight into a real jet would bankrupt even the Pentagon. An F-16 flight hour costs tens of thousands of dollars, Bill Gray, the Air Force Test Pilot School chief test pilot, pointed out recently. Neural networks need millions of training runs, and you simply cannot do that on actual airframes.

Instead, the Air Force matured the software across millions of hours in high-fidelity simulations. In just 24 hours, test pilot Maj. Ross Elder noted, the AI could train millions of times on rare edge cases seen only once or twice in real life.

Once the algorithms survived the virtual arena, they were tested onboard the X-62A, a heavily modified F-16D, across 10 validation flights. After further hardware-in-the-loop ground trials, the code finally graduated to the controls of the Valkyrie.

The test builds on four years of work dating back to the Skyborg Vanguard effort first unveiled in 2019, which already saw autonomous software fly on Kratos UTAP-22 Mako and General Atomics Avenger drones. Now, the tech is feeding directly into the Collaborative Combat Aircraft (CCA) program.

The 1,000-drone fleet

This Valkyrie flight is not just an isolated science fair project. It is the tactical blueprint for how the Air Force plans to fight future wars.

The service is preparing to buy at least 1,000 CCA drones. The operational plan envisions pairing two autonomous uncrewed wingmen with each of 200 next-generation NGAD fighters and 300 F-35As. These robotic sidekicks will scout ahead, jam enemy radars, or carry extra missiles.

Even so, military officials made sure to temper rogue-AI fears. The Air Force reiterated that for the foreseeable future, a human operator will remain firmly in the loop whenever these autonomous platforms are sent into action.

Why it matters

In our view, the most impressive part of this test is not the three-hour flight itself, but the pipeline behind it.

By leaning entirely into high-fidelity simulation and using testbeds like the X-62A to bridge the gap into reality, the Air Force has built a remarkably cost-effective way to train and deploy complex neural networks. Software that would previously take decades to validate in hardware can now be iterated overnight in a digital sandbox.

For the broader aerospace industry, this shifts the entire defense calculus. Future air dominance will not belong solely to the nation that builds the most expensive sixth-gen airframe, but to whoever can field mass quantities of affordable, autonomous drones to fly alongside them. With China's military aggressively pursuing tactical AI for its own fighter fleet, the race to operationalize uncrewed combat swarms is officially moving from the lab to the flight line.

The unresolved question is trust. The Pentagon insists humans will stay in the loop, but in supersonic combat where microsecond decisions dictate survival, human reaction time is inherently a bottleneck. Balancing genuine command oversight against the raw reaction speed of machine-learning models will remain the ultimate design challenge for autonomous warfare.