
What happened
AI agents built a first-person 'spot the anomaly' horror game in Unity, and the separate quality-check agent signed off as shippable, but a person playing 15 minutes found wall-clipping cameras, characters stuck inside desks and chairs, and said the game was not fun.
Why it matters
The AI reviewer only read code, scripts and logs, so it caught progression-blocking bugs but not what the screen actually looks like or whether players can notice and enjoy the anomalies — the game's core loop.
Summaries like this, every morning.
The account is a post-mortem of a workflow where the author split game development across specialized AI agents — programming, art, sound, scenario and quality checking — and ran several horror games in parallel. The quality-check agent worked as designed: it read code for dead ends, checked save data integrity, looked for contradictions in scripts and data, and ran automated tests, and it did find and fix several progression-blocking bugs. The failure was in what the review could see. Object placement lives as numbers in scene data, so overlapping a desk and a person is invisible unless you calculate the overlap or look at the rendered screen, and the AI reviewer never ran the game before declaring it shippable.
The deeper problem was not the visible glitches but the verdict that the game was not fun. For an anomaly-spotting game, if players cannot notice the anomalies, the core loop does not work, and fixing every collision would not change that. The author traces this to three limits: AI review can only judge readable files, not what appears on screen when the game runs; fun exists only in the person playing, and a separate attempt to have another AI (Codex) score the game out of 100 was an estimate from scripts and code, with the caveat that scariness and fun might differ from actual play; and the build order put review after everything was made, so quality rose but 'is it fun' was discovered only at the end.
The proposed fixes are procedural. Decide a one-line selling point and scenario first, build a small slice (one room, three anomalies) and have a person play for five minutes before building the rest, and make visible-breakage checks mandatory by instructing AI to run the game rather than read code — for example, a Unity editor test that computes collider overlap deeper than 2cm, Playwright to drive a browser build and capture screenshots at set points, and Unity play-mode tests that capture the same viewpoints on every fix. The author also notes that splitting work by role left no agent responsible for the overall experience, so a human director role is needed, and that ship criteria must be set before development begins. Time saved by AI, the author concludes, is better spent on early human playtesting rather than late review.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
MomongaCode is out under the MIT license, a terminal AI coding agent that works with Ollama, LM Studio, and Ap…

An FA engineer's Claude Desktop skill post-processes pptx output to remove per-character font direct specifica…

A developer's AI-operated cloud browser was blocked with HTTP 403 three times in one day when posting YouTube…

In a 2026-06-19 commit (b5af2be32), Shelly added a flow where a natural-language request is parsed by parseAge…

Ziva.sh tested GPT-6 Luna on 10 beginner Godot prompts, with 'Godot 4' specified, no version, and web search;…

From November 22, 2026, Gomumari's target-list AI agent will read a customer's past won-agency list — around 3…
