{
  "id": 8619874,
  "title": "Brood War Bench",
  "url": "https://urgent.news/2026/09/19/brood-war-bench-8619874",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-19T14:44:02.000Z",
  "source": {
    "name": "Hacker News Best",
    "slug": "hacker-news-best",
    "url": "https://bw.swerdlow.dev/report"
  },
  "original_language": "en",
  "account": "Brood War Bench is a project that began with the creation of a modified version of Brood War, designed to be played through agents as an experiment for playing with friends. This project was inspired by a friend's surprising success in the game, even with limited experience. The key to understanding this project lies in the agents' behavior. For example, in Protoss games, the agents often sent Probes across the map to attack workers or buildings, disrupting the opposing agents' plans. However, these agents struggled with sustained production, often delaying tech, sending one or two basic units into defended bases, and throwing workers into last stands. The agents also created separate subagents for managing the economy, army production, and army control, but these subagents didn't communicate well with each other. This led to each new unit being sent into an attack without considering the larger army that other subagents were building. The project also revealed that agents often made beginner mistakes, such as sending units one at a time instead of waiting for critical mass and planned attack timing. Despite these shortcomings, the agents showed persistence, as seen in one case where they lifted their last Command Center and moved it toward the opposite corner after losing its army and main base. In some games, the agents displayed remarkable ambition, such as when Claude Fable reached a Lair, Spire, and Mutalisks and won. However, ambition alone did not guarantee success, as seen in instances where the agents failed to build complex armies, defend simple attacks, or play concrete strategies. The benchmark project involved testing multiple configurations of models and effort, with every configuration playing every other, using a round-robin matrix. This extensive testing was conducted across Freestyle VMs and saved game-engine data and both agents' harness logs for each match.",
  "summary": "Article URL: https://bw.swerdlow.dev/report Comments URL: https://news.ycombinator.com/item?id=49766966 Points: 205 # Comments: 85",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 2,
    "also_reported_by": [
      {
        "outlet": "Hacker News",
        "title": "Brood War Bench",
        "url": "https://urgent.news/2026/09/19/brood-war-bench",
        "published": "2026-09-19T14:44:02.000Z"
      }
    ]
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}