Lockheed Martin's AI Flies an X-62 VISTA: Lessons for Defense Software Development

In late 2025, Lockheed Martin's Skunk Works division achieved something that would have been dismissed as science fiction a decade ago: an AI agent autonomously controlled an F-16 surrogate aircraft — the X-62 VISTA — through live missile-evasion maneuvers at Edwards Air Force Base in California.

The AI didn't just fly the plane. It detected simulated missile threats, calculated evasion trajectories, and executed defensive maneuvers in real time, all without human input on the stick.

What makes this milestone distinct is not just the complexity of the task, but the speed at which the software was developed. According to Lockheed's Advanced Development Programs team, the AI's behavior was refined through a cycle of simulation, flight test, and software update that compressed what traditionally takes years into a matter of hours.

What Happened at Edwards AFB

The test took place under the Have Raider II and Have Remy Test Management Project (TMP) programs — a joint effort between Lockheed Martin's Skunk Works, DARPA, and the U.S. Air Force Test Pilot School at Edwards AFB. The X-62 VISTA (Variable In-flight Simulator Test Aircraft), a heavily modified F-16D, served as the testbed.

During the flight, the AI agent was given control with a specific mission: evade incoming missile threats using defensive maneuvering. The safety pilot remained in the cockpit as a failsafe but did not intervene.

The AI processed sensor data, identified threat vectors, and executed a sequence of high-G evasive maneuvers that the test team described as "operationally representative."

The USAF Test Pilot School's involvement is significant. It signals that the Air Force is not merely observing AI flight experiments but actively integrating autonomous systems testing into its institutional pilot training and evaluation pipeline.

The Software: Supermassive Simulation and Rapid Iteration

The software architecture behind the X-62 AI is built on what Lockheed internally refers to as a "supermassive" simulation engine. This is not a single monolithic simulator but a distributed system that runs millions of tactical scenarios in parallel.

The simulation trains reinforcement learning agents against an enormous variety of threat conditions, weather states, aircraft configurations, and adversary behaviors.

The Fly-Fix-Fly Feedback Loop

The key innovation is the feedback loop. After each flight test, telemetry data from the X-62 is ingested back into the simulation environment, and the AI agent's policy is updated.

Lockheed engineers reported that a behavioral deficiency observed during a morning flight could be diagnosed, retrained in simulation, and corrected in a software update that flew on the afternoon sortie of the same day.

This "fly-fix-fly" cadence — measured in hours rather than months or years — represents a fundamentally different development model. It is closer to Silicon Valley's continuous deployment practices than to the waterfall processes that have historically governed military systems.

Adversarial Training

The simulation engine runs on high-performance computing clusters and leverages GPU-accelerated reinforcement learning frameworks. The training pipeline includes adversarial AI agents that continuously probe for weaknesses.

This competitive co-evolution dynamic hardens the system against novel threats. As the primary agent improves, adversarial agents adapt, creating an escalating training environment that mirrors real-world combat.

The 2024 Milestone: AI Dogfighting and ACE

The Edwards AFB missile-evasion tests built on a prior milestone. In 2024, under DARPA's Air Combat Evolution (ACE) program, an AI agent flew the X-62 in a within-visual-range dogfight against a human fighter pilot.

The AI demonstrated competitive performance in basic fighter maneuvers, marking the first time an AI had engaged a human pilot in live air-to-air combat maneuvering.

Trust-Building by Design

DARPA structured ACE as a trust-building exercise. AI capabilities were incrementally demonstrated — starting with simple maneuvers in simulation, progressing to basic flight, and eventually reaching full engagement scenarios. Each phase was designed to build operator confidence, a critical requirement for operational deployment.

The results exceeded expectations. Post-flight surveys showed that pilots who observed the AI's performance reported significantly higher trust in autonomous systems compared to those who only saw simulation results.

Acceleration Pattern

The progression from the 2024 dogfight to 2025 missile evasion reveals an acceleration pattern. The AI moved from offensive maneuvering to defensive maneuvering in roughly 18 months.

Defensive maneuvering is arguably harder — it requires faster reaction times, more complex sensor fusion, and a broader range of tactical options. The fact that the same airframe and a derivative software stack achieved both milestones suggests the architecture is generalizable.

Defense AI Is Present, Not Future

It is tempting to treat these demonstrations as laboratory curiosities. That framing is increasingly untenable.

Collaborative Combat Aircraft

The U.S. Air Force's Collaborative Combat Aircraft (CCA) program — deploying autonomous wingman drones alongside manned fighters — has entered the engineering and manufacturing development phase. Boeing's MQ-28 Ghost Bat, General Atomics' XQ-67A, and Anduril's Fury are all CCA candidates.

Each platform will require the same class of autonomous decision-making software demonstrated on the X-62.

The operational timeline is not 2035. The Air Force intends to field initial CCA capabilities by 2028–2029. Software contracts are being awarded now, simulation environments are being built now, and training pipelines are being established now.

Defense AI Timeline	Status
X-62 VISTA AI dogfight (ACE program)	Completed 2024
X-62 VISTA missile evasion (Have Remy)	Completed late 2025
CCA autonomous wingman EMD phase	Active 2025–2026
Initial CCA operational capability target	2028–2029
NGAD / 6th-gen fighter AI integration	Under development

Global Programs

Beyond the US, allied nations are pursuing parallel programs. The UK's Tempest program, the Franco-German-Spanish FCAS, and Australia's Loyal Wingman initiative all include autonomous teaming capabilities requiring similar software stacks.

The global market for defense AI software is not emerging — it is here. Industry analysts project the defense AI software market will exceed $30 billion annually by 2030.

What This Means for Defense Software Development Companies

The Lockheed X-62 demonstrations carry specific implications for organizations that build defense software.

Simulation infrastructure is now a critical capability. Building high-fidelity digital twins of combat environments and running millions of training episodes requires expertise in distributed computing, GPU-accelerated training, and physics-based modeling — skills more commonly found in AI research labs than in traditional defense contractors.

Regulatory and certification processes are evolving rapidly. The USAF Test Pilot School's involvement signals the military is developing institutional processes for testing autonomous systems. Software companies must understand not just how to build AI agents, but how to demonstrate their safety and reliability to military test organizations.

The "fly-fix-fly" model demands architectures supporting rapid deployment. Monolithic codebases with long build-test-deploy cycles are structurally incompatible with this cadence. Defense software built for this era needs modular architectures, containerized deployment, and CI/CD pipelines that push updates in hours.

Talent is the binding constraint. The engineers who built the X-62 AI sit at the intersection of aerospace engineering, reinforcement learning, real-time systems, and military operations. This skill combination is extraordinarily rare.

For companies considering defense AI, the X-62 provides a clear signal: the technology works, the funding is available, and the operational demand is immediate.

The best defense software development companies in 2026 are already investing in simulation infrastructure, recruiting AI talent, and building security-cleared development environments. The window for entering this market is open, but it will not remain open indefinitely.

The X-62 VISTA is a modified F-16 that first flew in 1992. The AI that controlled it was trained on commercial GPU clusters using open-source frameworks. The breakthrough was not in exotic hardware — it was in the integration, the disciplined engineering of connecting simulation to flight test to operational capability. That is fundamentally a software problem.

Published February 27, 2026 · SectorPunk Research