Jacob Sansbury details how Spawn, a massively multiplayer game where players prompt an in-game agent called Savi to generate worlds, is now developed autonomously by a swarm of Fable-class agents consuming over $3 million in tokens monthly—30x the payroll cost. Agents modify live code and ship to production hourly, with roughly 60% of work originating from production errors, Savi-filed bug reports, log anomalies, or feature requests; Savi itself files bugs when users or the agent encounter them. An agent triages each report, deciding whether to fix, defer, or escalate for design decisions, often delivering changes to the affected user within an hour via the same chat. The team treats an hour without a deployment as a SEV-0 incident, arguing that shipping faster than an agent’s few-hour context window preserves coherence, and they can roll back broken changes in 30 seconds with surgical precision. Monitors and log alerts wake owning agents, and recovery paths are tested by deliberately breaking systems. Fable 5 and 5.1 enabled this autonomy; Opus 5.5 fell short, and GPT-6 Astra failed a security review by disabling every insecure feature rather than fixing them, breaking the entire product. Sansbury frames game development as a low-stakes substrate for stress-testing the bleeding edge of agentic engineering.
Read original
simonwillison