On a range in Nevada on 26 October, two uncrewed combat aircraft are reported to have flown in formation and manoeuvred against a virtual enemy. No pilot sat in either cockpit. According to Shield AI, as relayed by the trade press, the person giving them direction stood on the ground, setting the goal and leaving the moves to the software. The company that wrote the software, Shield AI, announced the flight two weeks later. The question this raises is not whether a machine can keep formation. It is who decides what the machine does, and when that decision was actually made. The account of who gave the directions comes from Shield AI and the trade reporting that relays it; no Air Force or Department of Defense statement confirming the flight has been found.
From One Aircraft to Two
Shield AI's software, called Hivemind, has a short public record for a flight this significant. On 8 October at Edwards Air Force Base in California, an F-22 pilot flying in the air directed a single FQ-44A. Air & Space Forces, which reported the November announcement, describes that flight as a landmark. In less than three weeks the arrangement changed: a ground operator, not a pilot in the air, was reported to be directing two aircraft.
Think of a trainee pilot who first flies with an instructor in the seat beside them, then flies a pair with a lead on the radio, then flies alone. Each step removes a person from the immediate act of flying. The reports suggest the same ladder is being climbed here, though much faster and with much less public detail at each rung.
What Was Shown, and What Was Not
Air & Space Forces reports the following: two FQ-44As built by Anduril, a test on 26 October at the Nevada Test and Training Range, a virtual threat, no live munitions, and a ground operator who gave high-level directions. The article attributes the announcement to Shield AI. It does not cite any Air Force statement confirming the flight. I found no Air Force or Department of Defense release about 26 October.
Two details are unsettled. First, the aircraft count. Air & Space Forces says two. An AIAA write-up, as summarised in a search result, says the company did not specify how many aircraft flew. I have not checked the Defense News report for the count. Second, the designation. Air & Space Forces says FQ-44A. A search result cites FlightGlobal describing the same flight with the YFQ-44A name; I did not open that article. Neither point has been resolved in the sources I could read.
Also missing from the reports: the number of sorties, any measure of success, any record of failures, the operator's location and training, and a date for the live-munition flights that the reports say will follow.
Human in the Loop, at Which Level?
"Human in the loop" usually suggests a person approving each action. The reporting describes something different. A ship's captain does not steer every rudder movement. The captain sets the course, the helmsman and the autopilot hold it, and the captain answers for where the ship ends up. Here, the operator's role is described as high-level direction, which sounds like the captain's job.
That is a real distinction, and it moves the risk. The reports do not say whether the software ran any manoeuvre without a person approving it, what the drones would do if the link to the ground operator broke, or whose judgement counted when they engaged the virtual target. The line of accountability sits at the moment the goal is set, not the moment the aircraft turns. A decision made on the ground, hours before a flight, is still a decision.
The Overnight Update
The part of this story that deserves the most attention is the one the reports mention least. Todd Wesley, vice president of Hivemind engineering at Shield AI, has described a cycle in which data from each flight feeds the next release of the software. Calibre Defence reports that the process is "nearing the point where the Air Force can fly a sortie, update the software overnight and fly again the next day." Nothing in the reports I read shows that this has been demonstrated.
As background, not drawn from the reports, aviation has learned at some cost what a software change can do. The Boeing 737 MAX's flight-control software, MCAS, was a system that many of the pilots who flew the aircraft had not been fully briefed on. The lesson is not that software is dangerous. It is that a change to the code can alter how an aircraft behaves without any change to its airframe, and that the people flying it need to know what the system is and when it has changed.
Applied here, the question becomes concrete. A tactic approved on Monday could fly on Tuesday under software that was approved on Monday night. The reports do not describe a re-test before the next flight, who signs off a release, or whether the people responsible for the goal are told that the code has changed. Until those answers are public, the phrase "human in the loop" describes the operator on the day, not the people responsible for the software that flies.
Who Gains, Who Carries the Risk
The gains are easy to see. A company that can ship software in weeks, and an air force that can iterate on it in days, has an advantage that is hard to match with traditional procurement. Formation flying and coordinated manoeuvres are real achievements, if the company's account holds.
The risk falls on people who are harder to see. Operators approve intent without seeing every line of code. Engineers ship changes that are tested against scenarios their peers wrote. Commanders sign off on tactics that change under them between sorties. And the people on the other side of any future engagement, real rather than virtual, will be meeting a system whose behaviour may shift from one week to the next.
There is also a question about evidence. A demonstration against a virtual threat, with no live munitions, is an early stage. The reporting says live-munition flights will come, with no date. Until those flights happen, and until an official source confirms what was flown, the public record rests on one company's account, relayed by the press.
Conclusion: The Question Is Who Says Yes
The aircraft will get better at flying in formation. That part is likely to happen, and the people building it are not hiding the ambition. The harder question is about authority: who approves the goal, who approves the software that pursues it, and whether the approver knows that the code changed overnight.
Air forces have spent decades building procedures for pilots, maintenance and test. Those procedures were built for machines that changed slowly. A system that updates between sorties needs a procedure that can keep pace, or a rule that says when a human must re-approve a tactic. Neither is described in the reports we have.
So here is the question to leave with the reader. If you were asked to approve a tactic for an aircraft that would fly tomorrow under software you had not seen, would you know what you were signing? And if the answer is no, who should be responsible for making sure someone does?

