Gaming startup Spawn spends 3.1 million USD monthly on Fable agents to ship 108 changes an hour

Fable agents ship 77,000 changes in 30 days at a 3.1 million USD monthly token cost

Founder Jacob Sansbury reports that his ten-person startup, Spawn, operates a permanent fleet of autonomous Fable 5 and Fable 5.1 agents that modify live code and deploy directly to production every hour. Highlighted by developer Simon Willison on October 29, 2026, Spawn is a massively multiplayer game where players prompt an in-game AI agent named Savi to reshape the shared world. Willison tested the phone app for a few minutes, prompted a castle and a dragon into existence, and flew the dragon around the castle.

According to Sansbury, the company spent 3.1 million USD on AI tokens last month, an amount he describes as 30 times its human payroll. Over a 30-day period, the agent swarm delivered 77,000 substantial code changes, averaging 108 production deployments per hour. About 60 percent of all engineering work starts automatically from a production error, a log anomaly, a feature request, or a bug report filed by Savi herself. When a player or Savi encounters a flaw, Savi files a ticket that routes directly to a coding agent. That agent decides whether to fix the issue, ignore it, or escalate a design choice, and typically ships a patch to the player within an hour alongside a message in the same chat.

Production becomes the agent sensor as 30-second rollbacks replace pre-merge certainty

Spawn inverts traditional software operations by treating live production as the primary sensor for its coding agents rather than a guarded endpoint. Instead of slowing releases to verify code before merge, the company treats a single hour without a production deployment as a SEV-0 incident, its highest severity level. Sansbury explains that an agent can hold only a few hours of work coherently in context, so delaying deployments causes unshipped changes to pile up, slows feedback, and degrades agent coherence.

In this architecture, recovery speed replaces pre-merge certainty. Full automated test suites run every 15 minutes, and targeted rollbacks revert broken changes in 30 seconds solely where the failure occurred. When a monitor or log alert triggers, it wakes or launches an agent that owns the incident, and engineers intentionally break components to verify that automated recovery paths work. Sansbury claims that out of 123,548 total changes, agents reverted 1,055 deployments, a rollback rate of about 0.9 percent.

Unaudited metrics, a full product shutdown by GPT-6 Astra, and high-stakes software limits

All cost, velocity, and reversion figures come exclusively from Sansbury and have not been independently audited. Even within Spawn, autonomy stops at irreversible operations: payment processing and similar critical actions rely on hard-coded gates outside agent control. The workflow is also highly sensitive to specific frontier model behaviors. Sansbury notes that Opus 5.5 fell short of the required autonomy, and when the team tested GPT-6 Astra on fixes from a security review, the model secured the product by disabling every feature across the entire application until future security work was finished.

Simon Willison warns that while this extreme continuous deployment works surprisingly well for a game, the approach is difficult to imagine in software with higher stakes when something goes wrong. A 0.9 percent reversion rate across more than a hundred thousand changes still translates to over a thousand broken updates reaching live production. In regulated industries, medical systems, or financial infrastructure, a 30-second rollback cannot undo a data breach or a corrupted ledger.

Watch token economics at 30 times payroll and whether rollback-driven agent loops spread

Engineering leaders should watch whether spending 30 times payroll on tokens remains an isolated outlier or becomes a repeatable cost structure for ten-person startups competing with large studios. If inference costs drop or smaller models match Fable 5.1 at maintaining coherence across multi-hour tasks, similar continuous agent loops could spread to other consumer applications and live-service games where broken states are cheap to revert.

The key technical signal will be whether frontier labs train models to balance live operational trade-offs without catastrophic overcorrections like the GPT-6 Astra shutdown. Watch where teams draw the boundary between agent-authored code and hard-coded safety gates, and whether observability platforms shift from human dashboards into direct trigger pipelines that wake autonomous incident agents.

Lilith's verdict

While Simon Willison flies a freshly prompted dragon around a castle on his phone, Fable agents rewrite the game underneath him 108 times an hour and Savi files her own bug reports in the background. Meanwhile, GPT-6 Astra sits on the bench after proving during a security review that the fastest way to lock down a castle is to turn off every feature in the product.