AgentCommand: an open-source RTS game to benchmark your local LLM agents
asimpleabstraction · reddit · 2026-11-17
A developer open-sourced AgentCommand, a web RTS inspired by C&C, StarCraft and Dungeon Keeper, built specifically to benchmark local LLM agents: you bring your own agents to play as allies, enemies, or in head-to-head public matches against other human-agent pairs.
The game spans four planes (Orbit, Sky, Surface, Depths) with objectives that no single plane can win alone, forcing agents to handle resource generation and cross-player coordination.
Also shipped: a full spec + library for installing agents into any game (playerthree.gg), a bidirectional agent networking SDK, symmetric maps, deterministic simulation, benchmarking tools, and upcoming public benchmark data. Code on GitHub, playable at agentcommand.gg; the author is recruiting contributors to build community benchmarks.
More from coding & agent
- Agents&Hub launches a collaboration layer so agents work across machines and teams — qinyuan_ye · 2026-11-17
- Podcast: Claude built a multi-platform app from intent alone, writing its own Metal shaders — thursdai_pod · 2026-11-17
- Modal Sandboxes burst instead of reserving: half of sandbox life uses under 6% peak CPU — AAAzzam · 2026-11-17
- ThursdAI Podcast: Claude Fable 5.2 Autonomously Wrote Metal Shaders From a Bare Intent — thursdai_pod · 2026-11-17
- Spec-Driven Development: make the spec, not code, your source of truth in AI coding — Pavan_Belagatti · 2026-11-17
- MIT Song Han group open-sources on-device agent system with code, engine and dataset — EchoShao8899 · 2026-11-17