跪拜 Guibai
← All articles
AI Programming · Programmers

Epic Confirms Claude and Codex Write Unreal Engine Code: Ten-Year Bug Class Wiped Out in Months

By 深蓝AI ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Epic's confirmation ends the production-readiness debate: AI is now writing ship-grade Unreal Engine code, and a ten-year bug class fell in months. The pressure shifts to engineering processes, such as change-size limits, human review, and automated monitoring, that must keep pace with AI's far faster output speed.

Summary

Epic officially confirmed that Claude and Codex are writing Unreal Engine code. Marcus Wassmer, VP of Development, told GameDeveloper that AI tools now handle the bulk of code across three main streams: Fortnite, the UEFN editor, and next-gen UE6. The most concrete result: non-deterministic cooking defects, which had been unresolved for ten years, were swept away within months once AI tools were applied.

Epic also rebuilt its game launcher using vibe coding, making it noticeably faster, and uses AI to automatically monitor Fortnite's PC stability, proactively flagging problematic driver and hardware combos. Engineers no longer write most code by hand, but Epic controls scope through three boundaries: limiting AI changes to defined streams, keeping human review mandatory, and relying on automated observability for quick rollback.

The public debate about AI in production is effectively settled for Epic, but the controversy hasn't disappeared. Layoffs, Disney project delays, and sharp developer pushback on X sit alongside Sweeney's 'survive the AI revolution' framing. For teams adopting similar practices, the article's CI gate example translates those boundaries into enforcements: change-size limits, deterministic tests, and at least one human approval before merge.

Takeaways
— Epic's VP of Development Marcus Wassmer confirmed to GameDeveloper that Claude and Codex are used to write Unreal Engine code and features.
— Non-deterministic cooking defects, unresolved for ten years, were cleared within months once AI tools were applied.
— Epic's development now runs on three streams: Fortnite, UEFN editor, and next-gen UE6; a UE 5.9 may ship as a small update with AI-assisted changes.
— Epic rebuilt its game launcher through vibe coding, producing noticeably faster runtime.
— Epic uses AI to monitor Fortnite stability on PC, proactively flagging problematic driver and hardware combinations.
— AI use is constrained by three boundaries: limited scope to defined streams, human review of AI-generated code, and automated observability for quick detection and rollback.
— The CI gate example enforces change size (max 20 files, 800 lines), deterministic tests, and at least one human approval before merge.
Conclusions

Epic's willingness to name exact tools and specific use cases makes this a more credible data point than typical vendor AI claims, giving external teams a concrete pattern to replicate.

The CI gate design reflects a key consequence of AI acceleration: human review bandwidth becomes the bottleneck, so change-size limits are a direct, measurable response to AI's output speed.

The surrounding layoffs and Disney project delays suggest AI adoption in game development is driven by survival pressure as much as engineering curiosity; Sweeney explicitly frames the pivot as a way to survive the AI revolution.

Concepts & terms
Non-deterministic cooking defects
Defects in the asset cooking pipeline where identical input can produce different output between runs, often due to multithreading timing, cache state, or file system differences; they're notoriously hard to reproduce and locate.
Vibe coding
A development approach where a developer describes intent in natural language and AI produces runnable code; humans specify requirements and accept results rather than write code line by line.
UEFN (Unreal Editor for Fortnite)
A specialized Unreal Editor for building Fortnite creator content, one of Epic's three main development streams.
From the discussion
Featured comments
咬代码的兽

Epic directly using Claude to write UE6 is pretty intense. Now everyone is cramming agents into the core development flow; at this pace, an engine team's daily grind is going to be reviewing AI pull requests. I track all this new AI programming activity on ai345, it updates quickly.

See top comments, translated →
Source: juejin.cn ↗ Google Translate ↗ Backup ↗