Saab reveals its own combat AI agents after autonomous Gripen flight trials

Saab has moved its own air-combat AI from simulation into a fighter jet. In a November 3 announcement, the Swedish manufacturer revealed a family of in-house agents and said an agent autonomously controlled a Gripen E during combat scenarios. The intended roles extend from helping a cockpit pilot to operating in uncrewed aircraft.

From Helsing’s agent to Saab’s own software

The trials covered beyond-visual-range air combat: scenarios involving opponents outside a pilot’s visual range. Saab says that capability has now been validated under real flight conditions. Breaking Defense reports that testing and integration have predominantly focused on those scenarios, rather than the full range of tasks a fighter might perform.

Gripen has carried combat AI before. During a 2025 flight-test campaign, Helsing’s Centaur software took control of the fighter and autonomously performed beyond-visual-range maneuvers. The distinction in this announcement is ownership: Saab developed the new agents, their training environments and the supporting development tools internally.

Saab argues that this control lets it determine how the agents operate, how they are verified and how they evolve. Peter Nilsson, head of Advanced Programs in Saab’s aeronautics business, described the goal as networks of crewed and uncrewed systems working together, rather than individual aircraft acting alone.

The harder question is how we turn this success [of flying with the agents] into a dependable capability that can learn, improve, and be relevant over time

Peter Nilsson, Saab’s head of Advanced Programs, speaking to Breaking Defense and other media

Millions of simulated hours, then flight

The AI Fighter Pilot Academy is Saab’s mechanism for closing that gap. It develops the agents while building infrastructure to create, train and evaluate them across different scenarios and aircraft platforms. Its remit therefore extends beyond producing the software used in a particular flight trial.

Inside simulated environments, the agents accumulate millions of flight hours across multiple air-combat capabilities, according to Saab. Their actions are continuously evaluated against defined objectives. That process lets them explore tactics and decision paths, then adapt their behavior based on the results.

The company says it designed the infrastructure for reuse: new capabilities, scenarios and platforms can be added without rebuilding the development framework. Computing capacity comes in part from Sferical AI, a Swedish initiative in which Saab is a founding partner and which supplies large-scale computing for advanced AI development.

Nilsson also tied Swedish computing capacity to control over sensitive data. Using Google or Amazon instead could mean not having full control of that data, he told reporters. That makes the computing arrangement part of Saab’s stated case for retaining control of development, not simply a way to run more simulations.

An operational aircraft, not a finished combat capability

Gripen E supplies the bridge between training and flight. Saab says its software-defined architecture allows new AI software to be integrated into an operational fighter without a dedicated experimental aircraft. It also allows rapid testing of new technologies, including in civilian airspace, while keeping simulation and flight testing within the company’s development chain.

The announcement describes two intended settings for the agents, not just an autonomous replacement for a fighter pilot:

• Inside the cockpit, an agent can support the human pilot.
• On an uncrewed platform, an agent can operate alongside pilots flying other aircraft.

Saab says safety and control are built into development from the outset. But the announcement does not provide detailed safety-test results or combat-performance comparisons. Its stated next step is to broaden the agents’ capabilities and fly them in more scenarios—the continuing work behind Nilsson’s distinction between a successful flight and a dependable capability.