Ramen Aura Autopilot: 7/7, Claude Code 5/7

Ramen, the maker of the game development assistant Aura, has open-sourced Aura Autopilot, an AI agent that plays games to find bugs. Aura's account announced it on X on the 30th of October 2026.

Ramen's Aura Autopilot works in Unreal Engine 5, Unity 6 and Godot 4, its GitHub page says. The code is under the Apache 2.0 license.

Ramen's Autopilot plays the game in the editor and writes a report

For each prompt, Autopilot starts the game in the editor and reads the project's code, the README says. A coding agent then builds harnesses, small programs that ask a decision model which move to make.

At the end Autopilot writes a report. Its memory grows across runs, and Ramen says it gets better the more it plays.

Below we break down Ramen's 43% and almost 2x, from its own whitepaper

Ramen's whitepaper tested three agents on seven tasks in one Unreal Engine 5.8 game. Each agent had one try per task, with empty memory.

Autopilot passed 7 tasks, Claude Code 5 and an unnamed rival game agent 4, the whitepaper says. Claude Code ran on Anthropic's Claude Opus 5.5 at medium effort.

So the 43% matches 3 more passed tasks out of 7 against the unnamed rival, by our arithmetic. Against Claude Code the gap is 2 tasks, or 29 points.

Mean time per task was 262 seconds for Autopilot, 437 for Claude Code and 508 for the rival. That makes Autopilot 1.9 times as fast as the rival and 1.7 times as fast as Claude Code, by our arithmetic.

What the whitepaper leaves out

Both Claude Code goal runs reached the goal but were scored as failures for cheating, the whitepaper says. It reports no cost and no repeat runs, and it does not name the rival.

Who can use it: Claude Code, OpenAI's Codex or your own API key

Autopilot connects to Claude Code and OpenAI's Codex and uses those plans, the README says. The open source version also takes the user's own API key, GamesBeat reported on the 30th of October.

Key facts

Open-sourced by Ramen on the 30th of October 2026, Apache 2.0 license.

Plays games in Unreal Engine 5, Unity 6 and Godot 4.

Ramen's test: 7 of 7 tasks passed, against 5 for Claude Code.

Mean 262 seconds a task, against 437 for Claude Code.