Ramen open-sourced Aura Autopilot on October 30 for developers who still replay a level by hand after every change. The agent installs in Unreal, Unity, or Godot, reads the project, and plays in real time by injecting real user inputs, then reports whether a check passed. A task can be as narrow as confirming that hunger drops while the player stands still, or as long as crafting a stone hatchet and using it to kill a zombie. The Apache-2.0 repo covers Unreal Engine 5, Unity 6, and Godot 4 on Windows and macOS, and it runs on a Claude Code or Codex subscription or on your own API key. A coding agent researches the game and writes a harness of actions for that project. A decision model called Jev, through the TypeSafe SDK, then plays. Files kept per project, including a codex, notes, reusable harness code, and earlier traces, let the next run skip research it has already done. On Ramen's cold-start test of an unmodified Easy Survival RPG V5 in Unreal Engine 5.8, Autopilot passed 7 of 7 tasks at a mean of 262 seconds. Claude Code passed 5 of 7 at 437 seconds, and another game agent passed 4 of 7 at 508 seconds. Runs that cheated, by teleporting the player or spawning items, were marked failures. Those figures are the vendor's, from one game, with memory emptied before each run. Autopilot also comes with an Aura subscription, where Aura builds a gameplay feature and Autopilot checks it. Ramen, led by CEO and co-founder Andy Tsen, started in 2019 as the San Francisco VR studio behind Zenith: Nexus, then built agents that edit Blueprints, C++, and the editor. Nyx Iskandar and Tsen wrote the accompanying whitepaper. Outside that product, the plugin, harness runtime, and benchmark setup are now public on GitHub.