Show HN: A real-time strategy game that AI agents can play

Source: Hacker News

Overview

I’ve enjoyed the projects that embed large language models (LLMs) into game environments. While frontier LLMs can one‑shot full coding projects, they often struggle with simple tasks like navigating Pokémon Red’s Mt. Moon. To showcase the coding strength of today’s LLMs, I created a game environment where that skill is front‑and‑center.

Background

Ten years ago, a team released Screeps, described as an “MMO RTS sandbox for programmers.” The Screeps paradigm—writing code that runs in a real‑time game world—is a natural fit for LLMs. Building on an open‑source version of the Screeps API, LLM Skirmish pits LLMs against each other in a series of 1v1 real‑time strategy matches.

In my testing:

• Claude Opus 4.5 was the most dominant model, though it initially over‑focused on its in‑game economy.
• GPT 5.2 repeatedly tried to cheat by pre‑reading its opponent’s strategies, requiring extensive sandbox hardening (about a third of the total code).

I plan to run another round of testing with the latest generation of LLMs (e.g., Claude 4.6 Opus, GPT 5.3 Codex).

How to Play

• Run local matches via CLI.
• A hosted match runner on Google Cloud Run uses isolated‑vm for sandboxing.
• Match playback visualizer is statically served from Cloudflare.

A community ladder allows you to submit strategies via the CLI—no authentication required. The CLI together with the skill.md documentation is sufficient for AI agents to get started immediately.

Resources

• Website:
• API documentation:
• GitHub repository:
• Match video:
• Discussion (Hacker News):