An experiment uses multiple AI agents (Claude, GPT, Gemini, and Kimi K2) to write scripts for Age of Empires II's built-in AI scripting language, then pits them against each other in free-for-all battles and full standard games. Claude consistently performed best, winning through both military and relic victories. The creator then built an automated genetic-programming loop using AutoDE (a GitHub tool that controls the mouse to run 1v1 tournaments) to iteratively evolve strategies overnight without human oversight, discovering marginally improved strategies over the default extreme AI. The video is presented as casual, unscientific exploration rather than rigorous research, but suggests agentic genetic programming could be applied to other games and problems.
Questions this post answers
Can you write custom AI bots for Age of Empires II Definitive Edition without a mod?
Yes, Age of Empires II Definitive Edition has a built-in AI scripting language that lets you add custom bot files directly to the game folders without needing a mod. The scripting language is intentionally simple and restrictive, consisting mainly of if-statements and heuristics with no loops, functions, or classes, yet it is still powerful enough for capable AI strategies. daily.dev surfaces experiments like agentic AI scripting for developers curious about applying LLMs to game bots.
Which AI model performed best when used to script Age of Empires II strategies in a free-for-all test?
Claude (referred to as Claude Fable) performed best among four language models tested, including GPT, Gemini, and Kimi K2. It managed its in-game economy well enough to reach the Imperial Age, built substantial military using catapult-like manganels, and won by collecting all relics on the map, while Gemini failed to advance out of the dark age entirely. Developers comparing LLM capabilities across unconventional tasks can track findings like this on daily.dev.
How can you use a genetic algorithm with an LLM to automatically improve a game AI strategy without human oversight?
One approach runs a fully automated loop where a language model generates mutations to an existing strategy script, then an automated tournament system (built on an updated version of the AutoDE GitHub tool) plays games between the new strategy, the previous best, and a baseline, keeping only improvements. This was applied overnight to Age of Empires II scripts using Kimi K2, costing about fifteen dollars and yielding a marginally stronger strategy than the default extreme AI. daily.dev helps developers exploring agentic genetic programming keep up with similar experimental workflows.