A modified F-16 flew under the control of an artificial intelligence agent. VENOM is preparing the US Air Force’s future CCA combat drones.
In Summary
In July 2026, the US Air Force and DARPA flew a modified F-16 whose flight controls were operated by an artificial intelligence agent. A test pilot remained in the cockpit and was able to take back control at any moment. The event therefore represents neither the flight of a fully autonomous fighter nor the demonstration of an uncrewed firing capability. Its importance lies elsewhere. The VENOM program is converting several standard F-16s into testbeds capable of rapidly hosting different autonomy software packages. It builds on the work of the ACE program, which pitted the AI-controlled X-62A VISTA against a human-piloted F-16. This new phase is designed to prepare for beyond-visual-range combat, engagements involving multiple aircraft, and the control of future Collaborative Combat Aircraft. The American objective is clear: industrialize the development of combat AI and radically shorten the transition from simulation and experimentation to operational deployment.
The July Flight Validates a Testbed Rather Than an Autonomous Aircraft
On July 16, 2026, the Defense Advanced Research Projects Agency and the US Air Force announced that a modified F-16 had completed flight tests under the control of an artificial intelligence agent. Operations were conducted from Eglin Air Force Base in Florida.
Initial flights of the modified aircraft began in June to verify airworthiness, the operation of added equipment, and overall system safety. In July, the team moved to a more ambitious phase: autonomy software actually commanded the aircraft in flight.
This formulation must be understood precisely. The pilot had not left the cockpit. He monitored the agent’s behavior and retained the ability to interrupt its operation at any time. The US Air Force refers to this as a human-on-the-loop architecture. The human does not pilot every movement, but remains in a supervisory role with the capability to retake control.
The exact nature of the scenarios executed in July was not made public. DARPA did not specify the duration of AI control, the maneuvers executed, or the sensors utilized. No live firing, no autonomous lethal decisions, and no complete combat engagements against multiple adversaries were announced.
It would therefore be an overstatement to claim that an AI took sole command of a combat F-16. The test validates an experimental infrastructure, not an autonomous weapon system ready for deployment. Nevertheless, this infrastructure may have far-reaching consequences beyond the performance achieved during the flight itself.
The VENOM Program Converts F-16s into Flying Laboratories
VENOM stands for Viper Experimentation and Next-generation Operations Model – Autonomy Flying Testbed. The word Viper corresponds to the common nickname for the F-16 within the US Air Force.
The program aims to transform several existing aircraft into reusable testbeds. They are intended to allow for the rapid installation, testing, comparison, and modification of autonomy software developed by various entities.
This approach marks a departure from traditional aerospace development. Experimental software is no longer strictly bound to a single prototype. It can be integrated into an aircraft representative of an operational fleet, tested under real-world conditions, and then removed or replaced.
VENOM F-16s receive flight computers, wiring, software, measurement equipment, and specialized instrumentation. An automatic throttle system has also been added, allowing the AI agent to manage engine thrust in addition to the control surfaces governing roll, pitch, and yaw.
The artificial intelligence can thus command the aircraft’s entire flight path, modifying speed, altitude, orientation, and available energy—parameters that determine the effectiveness of a combat maneuver.
However, the most notable modification is less visible. The VENOM Autonomy Kit uses an interface connecting the artificial intelligence to the F-16’s flight controls, sensors, and mission systems. According to DARPA, this integration was achieved without rewriting the aircraft’s core software.
This separation protects the F-16’s critical functions. The artificial intelligence is not permitted to alter all onboard computers freely; it communicates with the aircraft through an intermediate layer designed to control its access.
The pilot can toggle between traditional controls and autonomous mode via a cockpit switch. This capability facilitates testing by allowing for direct comparison between agent behavior and pilot observations without grounding the aircraft for lengthy modifications between flights.
The AI Agent Is Not an Unbounded Digital Pilot
The phrase “artificial intelligence” can sometimes convey a misleading impression. The VENOM F-16 is not flown by a conversational assistant; it utilizes an agent specialized in tactical decision-making and control.
This software receives a representation of the operational picture, which may include the aircraft’s position, altitude, airspeed, fuel state, system status, and information regarding other aircraft. The agent then calculates an action aimed at achieving a defined objective.
For instance, it may seek to maintain an advantageous position, evade a threat, protect another aircraft, or vector toward an interception zone. It translates that decision into specific control inputs for the F-16.
DARPA has not published the detailed architecture of the agent used during the July flights. It is therefore impossible to state whether it relies exclusively on reinforcement learning, predictive models, or a combination of methods.
Another misconception should be set aside: the AI does not appear to learn freely in flight by rewriting its own rules. Data is recorded and analyzed after the mission, allowing developers to modify the model, test it in simulation, and prepare a new iteration.
This method prevents unpredictable behaviors from emerging unmonitored, creating a tight cycle of:
simulation, integration, flight, analysis, correction, and subsequent flight.
The speed of this loop represents one of VENOM’s primary objectives.
The ACE Program Demonstrated Within-Visual-Range Combat
VENOM does not start from scratch. It builds upon DARPA’s Air Combat Evolution (ACE) program.
ACE sought to develop an autonomous system capable of taking on a human-piloted aircraft in close-range dogfighting. This type of engagement represents a complex problem: both aircraft maneuver rapidly, constantly trade energy, and execute maneuvers under high gravitational forces.
Initial work was conducted in digital environments. Agents faced virtual opponents before competing against human pilots in simulators. DARPA subsequently transferred this software to the X-62A VISTA.
The X-62A is an F-16D extensively modified by the US Air Force Test Pilot School. Its flight control system can emulate the characteristics of different aircraft and host experimental algorithms without permanently altering the platform.
In 2023, an AI agent installed on the X-62A engaged in flight against a human-piloted F-16 in within-visual-range scenarios. DARPA made these tests public in April 2024.
That demonstration was significant, yet limited. It relied on a single experimental aircraft. While the X-62A is a unique flight laboratory, it cannot independently provide the flight volume required for the rapid development of multiple agents.
VENOM is designed to solve this bottleneck by translating the gains of ACE to a fleet of F-16s closer to standard operational aircraft. The goal is no longer merely to prove that an AI can maneuver, but to establish a continuous testing capability.
The AIR Program Prepares for Beyond-Visual-Range Engagements
Within-visual-range combat is not the primary mission of modern air forces. A vast majority of engagements are expected to occur beyond visual range.
In this environment, the pilot does not visually see the target, relying instead on radar, infrared sensors, airborne early warning aircraft, data links, and off-board information.
DARPA consequently launched the Artificial Intelligence Reinforcements (AIR) program. AIR builds upon the foundations of ACE but applies them to significantly more complex scenarios.
AIR aims to develop autonomy tailored for multi-aircraft engagements involving friendly and adversary formations. The agent must process incomplete information, anticipate enemy actions, and account for jamming, decoys, and tactics designed to deceive its sensors.
This evolution represents a fundamental shift. An agent operating in a close-range dogfight functions in a relatively readable environment, knowing the opponent’s position and directly observing its movements.
In beyond-visual-range combat, the situation is inherently uncertain. A radar track may be false, an opponent may turn off emissions, a data link may be jammed, and a missile launch may force multiple aircraft to break formation. The AI must make decisions without a complete tactical picture.
AIR targets two key technical advances:
First, building fast, accurate models of the environment, aircraft, and threats. These models must account for uncertainty and be continuously refined using test data.
Second, developing distributed autonomy, allowing each aircraft to make certain decisions locally while cooperating with the broader formation.
DARPA budget documents allocated $21.1 million for AIR in fiscal year 2024, rising to $41.2 million in 2025. This increase was intended to fund the transition from two-aircraft formations to four-ship scenarios, incorporating changing environmental conditions and electronic warfare capabilities.
Simulations Accelerate Learning Without Replacing Flight Tests
The development of combat artificial intelligence relies heavily on simulation. It would be prohibitively expensive and dangerous to test every new software iteration directly in flight.
The US Air Force indicates that a single scenario can be run up to 1,000 times in a digital environment. Engineers alter the initial positions of the aircraft, weapon parameters, adversary reactions, or detection conditions.
This repetition helps identify edge cases. An agent may succeed in 990 missions but fail in ten specific situations. Those ten failures can reveal a vulnerability that an adversary might exploit.
VENOM simulations began in 2024, initially focusing on one-versus-one engagements before expanding to two-versus-two scenarios at both short and long ranges.
The software is then transitioned to a software-in-the-loop environment, communicating with a digital representation of the aircraft’s systems.
The subsequent phase involves hardware-in-the-loop testing, where real flight processing hardware is connected to an F-16 simulator. Engineers verify response times, interfaces, and control safety features.
Safety limiters prevent the agent from exceeding the authorized flight envelope, ensuring it cannot command airspeeds, angles of attack, or load factors that could endanger the airframe or exceed the physiological limits of the onboard test pilot.
Flight testing remains essential. Digital models imperfectly replicate turbulence, vibrations, sensor latencies, measurement errors, and airframe dynamics. Operating in a real-world environment exposes the discrepancies between the model and reality.
The core value of VENOM lies precisely in this loop between simulation and flight test: real-world data refines the model, the model prepares the next iteration of the agent, and the rate of testing cycles increases.
VENOM F-16s Directly Pave the Way for Future CCAs
The ultimate goal is not to convert the entire US F-16 fleet into uncrewed aircraft. Rather, modifying these airframes serves to prepare for Collaborative Combat Aircraft (CCA).
CCAs are uncrewed combat aircraft designed to operate alongside crewed fighters such as the F-35, the future F-47, and potentially other platforms. They are not intended to function as simple remote-controlled drones.
Instead, they must execute portions of their mission with limited human supervision. Depending on their configuration, CCAs may carry sensors, electronic jammers, missiles, communications relays, or decoys.
A CCA could fly ahead of a crewed fighter to detect radar emissions, carry additional ordnance, jam adversary communications, or serve as a data link node between formations.
This distributed operational model offers significant utility in the Pacific theater, where distances are vast, bases are vulnerable, and crewed fighters are costly and available in finite numbers.
CCAs can increase the number of sensors and weapons in a given airspace without requiring an additional pilot for every airframe. They can also accept greater operational risk.
The challenge lies not merely in flying these platforms, but in enabling a pilot to assign them missions without micromanaging every maneuver.
The pilot must command, not remote-pilot.
AIR specifically aims to automate low-level tactical tasks. The human pilot sets objectives, priorities, and rules of engagement, while autonomous agents determine their own positioning, flight paths, and task allocation.
By July 2025, the US Air Force had demonstrated that an F-16C and an F-15E could each control two XQ-58A Valkyries during a training exercise. VENOM seeks to advance this concept further by developing more complex tactical behaviors that are less reliant on direct commands.
Open Architecture Prevents Vendor Lock-In
Combat autonomy cannot be developed under the traditional model of an airframe remaining functionally static for decades.
Algorithms will evolve faster than airframes, new sensors will emerge, adversary tactics will shift, and vulnerabilities will be discovered after exercises.
The US Air Force therefore seeks to decouple three distinct components: the aircraft, the baseline control system, and the mission autonomy software.
This approach relies heavily on the Autonomy Government Reference Architecture (A-GRA). A-GRA is a US government-owned framework that defines the interfaces required for different software packages to run across multiple compatible platforms.
The primary objective is to avoid vendor lock-in. Without an open architecture, the aircraft manufacturer would remain the sole entity capable of modifying its autonomy software, making every update dependent on a specific contract with that vendor.
Under A-GRA, the US Air Force aims to foster competition among software developers. An agent designed by a specialized software company could thus be tested on a platform produced by a different defense contractor.
In February 2026, the US Air Force indicated that RTX Collins was developing autonomy software for General Atomics’ YFQ-42, while Shield AI was working on Anduril’s YFQ-44. The tests were designed to demonstrate that the architecture could operate across different airframes and vendors.
VENOM complements this initiative by using F-16s as a common environment where multiple agents can be evaluated against identical scenarios. This enables the US Air Force to benchmark performance without building a new prototype for every software iteration.
The Approach Is Transforming US Military Acquisition
VENOM is more than an artificial intelligence initiative; it forms part of a broader transformation within US defense acquisition.
Traditional acquisition follows extended sequential phases: requirements are defined, a request for proposals is issued, a contractor builds a prototype, and flight testing begins years later—often leaving operators with a system whose foundational choices are already locked in.
The US Air Force is attempting to bring engineers, pilots, test units, and defense contractors together from the outset of a program.
At Eglin, the 40th Flight Test Squadron conducts developmental testing, evaluating technical functionality, safety, and baseline performance.
The 85th Test and Evaluation Squadron focuses on operational evaluation, studying tactical utility and how aircrews might employ the technology.
Because both squadrons operate from the same base, pilot observations can be fed directly to software developers, reducing the divide between technical testing and operational military evaluation.
A similar model is being applied to CCAs. In April 2026, an Experimental Operations Unit from Air Combat Command operated the YFQ-44A alongside the 412th Test Wing at Edwards Air Force Base. Operators participated in sorties, maintenance, and the development of initial procedures while the platform was still under development.
This acquisition framework accepts greater initial risk during early phases to field useful capability rapidly, improving it through iterative software blocks.
This approach is uniquely suited to software, where waiting for a complete solution risks rendering a system obsolete before it reaches operational deployment.
Trust Remains the Primary Operational Barrier
An artificial intelligence agent may outperform a human pilot in a defined scenario, but that does not guarantee an operator will trust it in actual combat.
Trust does not mean blindly relying on software; it requires that the human operator understands its limitations and can anticipate its behavior.
An overly cautious agent might abandon an advantageous tactical position, while an overly aggressive agent could consume fuel unnecessarily, expose itself to threats, or break formation. Software that excels in simulation may fail when a sensor feeds it anomalous data in the real world.
The challenge intensifies when multiple agents operate cooperatively, as a locally rational decision by one agent could disrupt the entire formation.
Artificial intelligence must also prove resilient against jamming and spoofing tactics. An adversary will attempt to generate false tracks, alter signatures, or provoke predictable automated reactions.
Engineers must ensure that agents do not depend excessively on a single data feed and that they maintain safe, predictable behaviors if communications are lost.
Cybersecurity is equally critical: compromised autonomy software could make flawed tactical decisions without causing an obvious system failure, making manipulation harder to detect than a mechanical malfunction.
Finally, rules of engagement remain a central concern. The July flight does not imply that an artificial intelligence is authorized to independently select and engage targets. US authorities emphasize that human supervision remains mandatory, though the precise boundaries applied in future operations will depend on mission parameters, communication links, and evolving doctrine.
The Real Test Will Be the Transition from Laboratory to War
The flight of the VENOM F-16 represents a genuine step forward, demonstrating that the US Air Force and DARPA can integrate an autonomous agent into a conventional combat aircraft without completely redesigning its underlying system architecture.
Yet the milestone should not be overstated. A pilot was present in the cockpit, control could be reclaimed at any time, and the tactical performance of the software remains classified. Complex multi-aircraft engagements, resistance to jamming, and autonomous weapons employment have not been publicly demonstrated.
Nevertheless, American ambitions are clear: Washington is no longer merely aiming to build an AI capable of winning an experimental duel, but is working to build a combat software production pipeline.
This pipeline links simulation, VENOM F-16 testbeds, flight testing, open architecture, and future CCAs. It is structured to evaluate multiple agents, port them across different platforms, and update them far more rapidly than traditional aerospace programs permit.
The next milestone will be considerably more challenging. Autonomous agents will need to operate in formation, utilize real-world sensors, function despite electronic jamming, and respond to adversaries whose actions have not been pre-programmed. They must also reduce pilot workload rather than burdening the crew with the constant monitoring of multiple uncrewed platforms.
Ultimately, VENOM’s success will not be measured by the number of F-16s capable of flying unassisted, but by the ability of a human crew to command multiple autonomous aircraft in a complex environment without losing control of the mission. At that point, human-machine teaming will cease to be a technological demonstration and become a operational airpower capability.