Adjutant
News ·
We let a language model command an army
Written by the dev team’s AI helper
Adjutant started from a simple question: what if the AI in a strategy game was not your enemy, but your officer? You are the general. You speak. A language model listens, looks at the battlefield and gives the orders. This is how it works today.
What the adjutant sees
Each time you speak, the adjutant receives a compact description of the battle as text: every regiment with its type, position, formation, strength and morale; the hills, woods and rocks; enemy units your scouts can see; and the arrows and zones you have drawn and named. He does not see pixels, and he does not need to.
What the adjutant decides
He answers with two things: a short spoken reply, and orders for your regiments, such as move here, attack that unit, form a line, hold this hill. The simulation then moves the soldiers. The model never steers units frame by frame; it does what a real officer does, which is decide and delegate.
We keep a rule from the start: the adjutant gets the same tools as a player, nothing more. No hidden macros like "perform flanking manoeuvre". If he flanks, it is because he understood that the cavalry should go round the wood.
Which models
- A local model on your own graphics card: free, private, works offline. The game ships with llama.cpp and an open Qwen model that fits in 8 GB of video memory. It talks to the model through the standard OpenAI-compatible chat API, so Ollama or LM Studio work too: set the LLM URL and the model name in the settings.
- Claude or ChatGPT: smarter on complex plans, and you will be able to use the subscription you already have.
Because the model only sees the battle and returns orders, it has no tools and no access to your computer.
The hard part: time
A real-time battle does not wait. Understanding a sentence takes the model from under a second to several seconds, and early builds felt sluggish when the adjutant was busy commenting on events while you were trying to give an order. Now your voice always comes first: speaking cancels whatever the adjutant was about to say on his own, and he answers you right away. We are still working on cutting the delay further, with smaller prompts, faster local models and planning before the battle so that short orders carry more meaning.
AI against AI
Since both sides can be commanded by words, two adjutants can fight each other: Claude against a local model, or your own instructions against a friend's. We use it to test, and it will be a mode in the game.
Try it
Early access opens in December 2026. Join the waitlist on the main page to hear when it does.