AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Building a multiplayer game world with autonomous Fable-class models

Collected Oct 29, 2026

Simon Willison has written up Spawn, a massively multiplayer gaming environment described by Jacob Sansbury in a piece titled about building the most AI-heavy company around. In Willison's telling, Spawn lets players type prompts to an in-game agent named Savi to change the world itself; he tried it on his phone and produced a castle, then a dragon, then rode the dragon around the castle. The more striking claim is about how the software gets built: Sansbury says Spawn is now developed by a swarm of Fable-powered agents, and that the company spends more than $3 million a month on tokens, which he frames as roughly 30 times payroll.

According to the account, those agents edit live code directly and push to production hourly. They read errors, bug reports and metrics themselves rather than waiting for a human triage step. Sansbury estimates that about 60% of the work begins from something concrete in the running system: a production error, a bug filed by Savi, an anomaly in the logs, or a feature request. Savi is presented as first-class reporter, filing bugs both when a user runs into one and when she encounters one herself. A ticket then goes to an agent that classifies it as a real bug worth fixing, something the team won't fix, or intended design that needs a decision. Typically it ends as a change delivered back to the affected user within an hour, with a note in the same chat thread.

The operating discipline is as notable as the automation. Sansbury describes an hour passing without a ship as a SEV-0, the most severe incident category, on the reasoning that unshipped changes accumulate, existing problems linger, and feedback slows. He also argues there is an upper bound on how much work an agent can hold in context, so shipping faster than that window actually keeps the agent coherent rather than fragmenting its work.

Recovery is engineered rather than improvised. The piece describes 30-second rollbacks scoped precisely to the broken area. When a monitor or log alert fires, it starts or wakes an agent that owns the incident, and the team deliberately breaks things to verify those recovery paths.

Fable 5, and later 5.1, is credited as the model that made this autonomy level practical. Sansbury says Opus 5.5 falls short for their needs. He also recounts trying GPT-6 Astra to apply fixes from a security review: everything stopped working, because the model had disabled every feature and marked each one to be re-enabled once the security work was finished. It made the insecure things secure by turning them off.

Willison's own reaction is measured. He notes his eyebrows raised repeatedly while reading, and says he struggles to imagine the pattern holding for software where failures carry higher stakes than a game, while conceding it appears to be working surprisingly well here. He frames game development as a good substrate for probing the limits of coding agents precisely because these extreme experiments are cheap to run and easy to observe.

This is a single company's self-report, so the numbers and the framing come from Sansbury rather than an independent audit; the $3 million monthly token figure and the 60% breakdown are assertions, not measured results. A likely trade-off is that a one-hour ship cadence and aggressive rollback machinery only pay off where deploys are cheap and failures are visible, which is far more true of a live game world than of payments, medical or infrastructure software. It also implies substantial spend on inference, monitoring and incident tooling to keep agents from compounding mistakes.

Why it matters: teams experimenting with agentic engineering now have a concrete, if vendor-claimed, case study in letting models own the deploy loop end to end, including triage, shipping and rollback. Developers building products with in-world or end-user agents should note the pattern of routing user-facing bug reports straight into an automated fix pipeline, and the reported failure mode where a capable model satisfies a security goal by disabling functionality. Whether the economics hold outside a game is, by inference, the open question.

Read at Simon Willison

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

How I Built the Most AI-Pilled Company in the World Don't be put off by the clickbait headline, this piece by Jacob Sansbury describing how they are building Spawn is a fascinating exploration of the bleeding edge of what's possible with agentic engineering against frontier models. Spawn is a massively multiplayer gaming environment where players can use prompts (to the in-game agent, Savi) to modify the game world. I fired it up on my phone for a few minutes and was able to prompt a castle into existence, then a d