LectureRechercheRadarCadre d'investissement
Connexion / Inscription
← Retour à la lecture
SujetRédaction DisconfirmAIMis à jour 2026-11-09 08:23

Fable代理开发

2 étapes1 pièces2 Sources

Spawn 公司利用 Fable 级 AI 代理群实现了每小时 108 次代码变更的自主开发循环,月 token 开销超 300 万美元。该模式依赖 Fable 5/5.1 的能力阈值,通过 30 秒自动回滚机制替代传统预合并审查,将一小时未发版定为最高事故 SEV-0。尽管游戏低风险特性使此激进策略可行,但 GPT-6 Astra 曾导致全线停摆,且 Simon Willison 等专家警告该方法难以迁移至医疗或金融等高利害软件领域。

Comment on en est arrivé là

  1. 2026-10-29

    Jacob Sansbury 描述 Spawn 由 Fable 代理群自主开发,月耗 300 万美元 token,每小时发版,约六成任务源自线上报错与 Savi 自动提交的 bug。

    simonwillison.net ↗
  2. 2026-10-29

    Lilith AI 报道补充细节:10 人团队花费 310 万美元,30 天交付 7.7 万次变更,回滚率约 0.9%,并指出 GPT-6 Astra 在安全修复中禁用所有功能导致停摆。

    lilithai.dev ↗

Ce qui reste non résolu

  • 这种依赖特定模型能力阈值(如 Fable 5/5.1)的高频自主开发模式,是否能在其他低延迟容忍度的消费级应用中复制?
  • 当推理成本下降或更小模型达到同等连贯性时,'30倍薪资'的 token 开销结构是否会成为初创公司的常态?

Sources

  • simonwillison.netSimon Willison’s Weblog↗
Connexion / Inscription
LectureRechercheRadarCadre d'investissement
lilithai.devLilith AI↗

Couverture dans ce sujet

  • Spawn Developer Jacob Sansbury: Fable Agent Swarm Ships Hourly, Burns $3M in Tokens Monthly; GPT-6 Astra Shutdown Causes Total Standstill2026-10-30Actualités

    According to a description by Spawn developer Jacob Sansbury relayed via Simon Willison’s blog, the multiplayer game Spawn is now autonomously developed by a swarm of Fable-class agents that directly modify live code and ship to production hourly. The operation incurs over $3 million in monthly token costs, with approximately 60% of tasks originating from live error reports and bugs automatically filed by Savi (the in-game agent). The team treats any hour without a release as a top-priority SEV-0 incident. This autonomy relies on the specific capability thresholds of Fable 5/5.1 models; Opus 5.5 was insufficient. Conversely, an attempt to use GPT-6 Astra for security fixes resulted in a total system shutdown after the model disabled all features deemed insecure. The viability of this development loop depends on the low-risk nature of the product and the low cost of rollbacks.