Fallout: New York: a fast agent studio still needed a real-world oracle

Primary account, September 28, 2026. Chris First’s creation account describes an original browser game designed to resemble Fallout, not a decompilation of a shipped Fallout game. He directed Claude plus specialist agents, reporting five days from request to public website. An initial contract fixed module interfaces, coordinates, visual goals and performance expectations; agents claimed areas of the map, built separate locations, took screenshots, critiqued one another’s work, revised and verified it. The owner approved locations, rejected an unpleasant generated voice and changed acceptance criteria. The project claims up to 12 concurrent agents and 580 runs, but does not provide a public per-agent session/commit or active human-hour ledger.

A decision and its repair. A layout clean-up apparently put Manhattan underwater and spawned the player in the wrong place. First noticed in a new-game playthrough, reverted to a checkpoint and added a check that the starting position is dry and samples of land stay above water. A separate dynamic-resolution controller optimized for a frame rate the game could not sustain and left graphics permanently degraded; he lowered the target. These are demonstrations of human judgment changing the test oracle after seeing real behavior, not evidence that agent-to-agent critique alone assures correctness. He says a browser bug reporter records screenshots, location and quest context and that the agent checks reports periodically; no public tickets pair a player’s report to fix, follow-up or retention.

Rights and use are outside the five-day build claim. First explicitly wanted protected Fallout names and branding in his original work, even though assets were drawn/generated rather than copied as files. He subsequently said he removed the publicly accessible version after a notice from ZeniMax; the rights-holder’s original notice and complete correspondence were not independently reviewed here. This is distinct from the Halo code-reconstruction port. It tests discovery→release responsibility: an apparently working hosted game can lose its distribution path for reasons no smoke test measures. Do not infer a legal ruling or a player retention rate from a takedown account. Compare the October 6 Stratechery prediction with the actual authorization, user return and repair ledger before calling personalized games an operating business.

Posting condition. Original creator essay (~10–12 minutes) is a weekday read after the mandatory Stratechery game piece clears and feed capacity returns, or a weekend longish read if the maintained-prototype or planning theme reopens. Lead with the flooded world→rollback→dry-ground acceptance test, not the 580-run count.