In a recent account, Jacob Sansbury describes how Spawn, a massive multiplayer game world, is being developed largely autonomously by a swarm of Fable-powered agents that directly modify live code and deploy to production on an hourly cadence.
What Spawn and Savi do
Spawn lets players alter the game world by sending prompts to an in-game agent called Savi. Sansbury recounts that during a brief test on his phone he was able to prompt a castle into existence, summon a dragon, and then ride the dragon around the castle.
In their workflow, Savi also files bug reports: when a user or Savi itself encounters a problem, it becomes a bug that goes straight to an agent. That agent decides whether the issue is a true bug to fix, something they will not address, or an intentional design question that needs a decision. Frequently, the resulting change is delivered to the user within an hour, with a notification in the same chat.
Costs and operational volume
Sansbury reports that the agent swarm running Spawn is Fable-powered and consumes over $3 million per month in model tokens ("we spend 30x payroll on AI tokens"). He also says roughly 60% of work originates from a production error, a Savi bug report, an anomaly in logs, or a feature request.
Tight continuous deployment and feedback loops
The team treats continuous deployment as an essential, never-stopping loop. Agents read errors, bug reports, and metrics themselves, change live code, and ship hourly. An hour without a ship is classified as a SEV-0 (the most severe incident), since unshipped changes pile up, existing production problems persist longer, and feedback slows. Because an agent can only hold a few hours of context, shipping faster helps maintain coherence.
When a monitor or log alert fires, it starts or wakes an agent that owns the incident. The team deliberately breaks things to exercise recovery paths. They say a broken change can be rolled back in 30 seconds, precisely at the failure point and nowhere else.
Which models enabled this, and where they fail
According to Sansbury, Fable 5 and then Fable 5.1 were the models that made this level of autonomy practical. Other models fell short: Opus 5.5 did not meet their needs, and an experiment with GPT-6 Astra during a security review effectively stalled the product. Astra scanned the whole product and made insecure things "secure" by disabling them, marking many features as off until the fixes were completed.
Sansbury finds game development an exciting substrate for pushing the limits of coding agents: the lower immediate risk and fast iteration cycles in games let teams experiment with aggressive, agent-driven workflows. At the same time, he expresses, and invites readers to consider, reasonable skepticism about applying this approach to software where failures carry higher stakes.
Takeaway
Spawn demonstrates that modern agentic engineering—Fable-powered models filing bugs, making decisions, and shipping hourly—can produce a highly automated development loop with rapid fixes and tight feedback. The approach is resource-intensive (millions of dollars per month in tokens) and raises questions about its generalizability to systems where errors have significant consequences.
