Twenty-seven times, the X-62A VISTA found a live T-38 Talon with its infrared sensor, calculated an intercept, and maneuvered to kill position. The pilot in the cockpit never touched the controls. The AI flying the jet was not made by Lockheed Martin. It was integrated in a single quarter.
On August 4, 2026, the U.S. Air Force and Lockheed Martin confirmed what test pilots at Edwards Air Force Base had known for months: a heavily modified F-16D, the X-62A VISTA, had used a third-party AI agent to autonomously intercept a crewed T-38 across eight flights, executing 27 successful engagements. The AI received a live feed from a Lockheed Martin Legion Pod, an infrared search-and-track (IRST) sensor, processed the targeting data, and generated flight commands in real time. A safety pilot sat in the cockpit, hands off the controls.
This was not a simulation. The target was a physical aircraft flown by a human. The sensor was operational hardware. The AI was not a prime contractor's product. It was a third-party autonomy agent integrated onto the jet in three months.
From dogfight to hunt
The X-62A is no stranger to AI. In 2023 and 2024, the same aircraft flew the DARPA Air Combat Evolution (ACE) program, the first in-air tests of AI algorithms autonomously flying an F-16 in within-visual-range dogfights against a human-piloted F-16. ACE proved an AI could maneuver and shoot. But those engagements relied on virtual sensors and controlled starting positions. They answered the question: can an AI fly?
HAVE HEAT answered a harder question: can an AI hunt? The distinction is fundamental. A dogfight assumes you have already found the enemy. An intercept requires finding the target first, often at beyond-visual-range distances, using onboard sensors degraded by weather, jamming, and clutter. The Legion Pod's infrared data streamed into the AI's decision loop, and the AI translated raw sensor pixels into a three-dimensional intercept geometry. According to the Air Force's 412th Test Wing, the agent "used targeting information from an operational sensor to execute successful air intercepts against a live target." The leap from ACE to HAVE HEAT is the leap from stick-and-rudder skill to sensor-fusion warfare.
HAVE HEAT was one of two rapid experiments under the X-62's Mission Systems Upgrade (MSU) program. The other, HAVE HOLIDAYS, focused on integrating a different non-prime autonomous agent and testing enhanced safety rules for advanced sensor exploitation. Together, they form a new testing paradigm: bring in a third-party autonomy vendor, wire them into an open-architecture mission computer, strap a new sensor to the aircraft, and fly combat-relevant missions. All in months.
Three months. Not five years.
The three-month timeline is the real story. The X-62A VISTA, assigned to the USAF Test Pilot School, is undergoing a deliberate transformation into a modular, open-architecture testbed. The MSU program strips out the proprietary, monolithic avionics that define most legacy fighters and replaces them with a flexible system designed to accept third-party software and new sensor packages as plug-and-play modules.
Contrast this with the standard acquisition cycle. Integrating a new sensor onto a frontline fighter like the F-35 typically takes years. The software changes alone trigger cascading safety certifications, prime contractor negotiations, and regression testing across the entire mission systems suite. The sensor becomes inseparable from the platform. Changing anything means changing everything.
The HAVE HEAT team took a different path. They integrated a non-prime autonomous agent, fed it live Legion Pod data, built the safety constraints for autonomous intercepts, and cleared the aircraft for flight in three months. Lt. Col. Joshua Strafaccia, dean of faculty for research at the Air Force Test Pilot School, described the result as a "meaningful expansion of avionics capability towards integrated, AI-driven control of multiple sensors and air vehicles for mission autonomy." The quote is careful. The implication is blunt: the Test Pilot School has built a pipeline that can field combat-relevant autonomy faster than the primes' traditional processes.
Code is no longer the constraint
The DARPA ACE program proved AI can dogfight. HAVE HEAT proved AI can use a live IRST sensor to autonomously intercept a crewed target. The algorithms are mature enough to fly and fight. The constraint is no longer code.
The new rate-limiting step is sensor-hardware integration speed. The question shifts from "can the AI do it?" to "how fast can you wire the AI into a real sensor and get it airborne?" This is a manufacturing and software-architecture problem, not an algorithmic one. Whoever masters that pipeline first will dictate the pace of unmanned air combat over the next five years.
This puts a specific class of winners and losers into sharp relief. The winners are primes and labs that have already committed to open-architecture mission systems. Lockheed Martin's Skunk Works, which built the X-62's backbone and is the primary integrator on the MSU program, has a structural advantage. Third-party autonomy vendors who can demonstrate repeatable, rapid sensor integration will find a hungry customer.
The losers are primes still defending proprietary, monolithic avionics stacks. A closed system that requires a five-year block upgrade to add a new sensor is incompatible with an operational tempo measured in weeks. The market signal from Edwards is clear: the Air Force will no longer wait for the platform owner to bless every new capability.
The CCA program just ran out of time
The second-order consequence is a forced acceleration of the Collaborative Combat Aircraft (CCA) program. The CCA initiative envisions fleets of autonomous drones flying alongside crewed fighters. The program's official timeline has pushed initial operational capability toward the late 2030s.
HAVE HEAT collapses that timeline. Here is the mechanism. The autonomy agent that flew 27 intercepts on an F-16 testbed can be transitioned directly into a CCA prototype. The sensor-integration pipeline developed under the MSU program provides the template. The Air Force now has empirical proof that a third-party AI, integrated on an open-architecture platform, can perform a combat-relevant mission against a live target. The program office can no longer argue that the autonomy is immature or that integration is too complex. The test data exists.
Within 12 to 18 months, expect the Air Force to announce a follow-on program that ports the HAVE HEAT autonomy and integration approach directly into a CCA demonstrator. First operational CCA deployment will likely move from the late 2030s to 2028 or 2029. The falsification condition is straightforward: if no such program is announced by early 2028, the acceleration thesis is wrong. But the bureaucratic logic is inexorable. The Test Pilot School has demonstrated a faster, cheaper path to combat autonomy. The CCA program office must either adopt it or explain why it cannot.
The Test Pilot School's open-architecture approach becomes the de facto standard by default, not by committee consensus. The institution that proves it can field combat autonomy in three months will have its methods copied. Future MSU upgrades, including an AESA radar and increased onboard computing, will only widen the gap between platforms that can absorb new autonomy and those that cannot.
Fix the pipeline
The operational takeaway is straightforward. Funding should shift from broad AI algorithm development toward rapid sensor integration and open-architecture mission systems. The Test Pilot School's model should expand and institutionalize across the combat air forces. The CCA program office should restructure its milestones to prioritize integration speed as a primary metric.
The HAVE HEAT tests were not a proof of AI capability. They were a proof of integration tempo. The 27 intercepts did not demonstrate that AI can fly and fight. DARPA already proved that. They demonstrated that a small team with an open-architecture aircraft and a clear mandate can field combat autonomy against a live target in three months.
The algorithm is ready. The question is: how fast can you wire it in?
