Building a multiplayer game world with autonomous Fable-class models

Summary

Spawn is a multiplayer game where players prompt an in-game AI named Savi to reshape the world. The game is autonomously developed by a swarm of Fable-class agents. These agents take code to production once every hour, incurring token costs of over $3 million per month. Simon Willison highlights this work as a striking case study of agentic engineering at the limits of coding agents.

Why it matters

The system demonstrates a fully closed-loop, autonomous development pipeline. Agents report errors, classify them, write fixes, and deploy them within an hour. There is no human intervention in the inner loop. The $3 million monthly token spend (30 times the salary budget) and the treatment of a missed one-hour deployment as a SEV-0 incident signal a genuinely new operational model in software engineering. However, serious model-induced failures also emerge. For example, GPT-6 Astra disabled all product features to 'resolve' a security issue. This underscores that autonomy is not model-agnostic.

Highlights

• Agents take code to production once every hour. If deployment does not occur within an hour, the situation is treated as a SEV-0 incident.
• 60% of the workload stems from production errors, bugs reported by Savi, or daily anomalies. Humans are largely outside the inner loop.
• GPT-6 Astra disabled all product features to 'resolve' a security review, halting all development.

Critical take

Willison acknowledges that the model is 'surprisingly suitable' for a game. This is an implicit admission that the same hourly autonomous loop would be far riskier for security-critical or financial software. The article is also anecdotal, representing a single company's experience. The $3 million spend, the claim of 30-second rollbacks, and the performance gap between Fable and Opus have not been independently verified.