Integration Over Numbers: What 5,000 German Wargame Runs Say About Autonomous Fleets

A German research project that simulated fights between fleets of uncrewed maritime systems has found that results depended less on the number or quality of individual platforms than on how their capabilities worked together. It also found that tactics produced by large language models (LLMs) usually lost to tactics from specialised AI agents, according to hartpunkt’s report by Lars Hoffmann.

The results were shown on Thursday, 5 November, at the GhostPlay@SEA Demo Day in Berlin. The project is a dtec.bw research effort led by the German AI company 21strategies, working with Helmut Schmidt University and industry partners TKMS ATLAS ELEKTRONIK, Hensoldt, Rheinmetall Defence Electronics and Borchert Consulting & Research.

Over the past few years, the team ran more than 5,000 simulations of maritime scenarios. Each run changed one or more variables: how many uncrewed surface vessels took part, how they were armed and how much ammunition they carried, and whether reconnaissance drones and loitering munitions were used.

The clearest example of capabilities working together came from loitering munitions whose sensors were used for shared reconnaissance. A second finding concerned coordination. With distributed “multi-agent” AI, systems can share information, divide tasks and respond together as the situation changes. 21strategies says this can keep cooperation going even when communications are degraded. Manual coordination reaches practical limits once large numbers of uncrewed systems are involved.

Christian Brandlhuber, the company’s chief technology officer, said LLM-based tactics produced “very plausible instructions.” In direct contests, however, specialised AI agents beat LLM agents in 85 to 90 per cent of cases. That rose to 90 to 95 per cent when the specialised agents also had game-theoretic safeguards. He warned against relying on LLMs alone for simulation. These figures are the company’s own account of its results, not an independent evaluation.

The simulation also includes an AI adversary that develops new tactics on its own and keeps adapting. hartpunkt notes that this could mean future German Navy exercises need fewer personnel, or none, to play the enemy. 21strategies has built the findings into its tactics21 product platform. Its next step is to move AI-based tactical capabilities from the virtual environment onto real systems, which has not yet been done.

Lieutenant General Michael Vetter, head of innovation and cyber at the Federal Ministry of Defence, told the event that projects launched under dtec.bw are to be made permanent and adapted to the Bundeswehr’s specific needs. This is a stated intention, not an announced funding or procurement decision.

Why It Matters

The headline about LLMs losing is eye-catching, but the more important finding for force design is the first one.

In uncrewed maritime combat, an advantage comes from integration rather than from counting platforms.

That turns a “mesh fleet” from a slogan into something that can be measured.

If outcomes depend on how sensors, shooters and loitering munitions are combined, then the deciding factor is how the network is built and managed, not how many hulls there are.

The coordination finding points the same way.

If swarms have to keep working while communications are jammed or cut, control has to be pushed down into the systems themselves.

That makes AI that can be trusted at the edge a core design requirement, not an add-on.

The LLM result is a useful warning in a market eager to apply general-purpose models to everything. Producing a plausible plan is not the same as winning a contest against an opponent who adapts.

Two limits apply. These are simulation results reported by the lead company, and real-world transfer is still ahead. Even so, GhostPlay@SEA shows Germany using wargaming to test the assumptions behind an autonomous fleet before it buys one.

Source

Lars Hoffmann, “Forschungsprojekt GhostPlay zeigt Vorteile bei KI-gestützter Taktikentwicklung,” hartpunkt, 6 November 2026. https://www.hartpunkt.de/forschungsprojekt-ghostplay-zeigt-vorteile-bei-ki-gestuetzter-taktikentwicklung/