What if you could give an AI agent your unfinished game and tell it:
"Play this. Find everything that's broken. And don't stop until you've tested it properly."
Well, someone built exactly that.
Meet Aura Autopilot, a new open-source AI game testing agent that actually plays your game, investigates the code, discovers bugs, and gets better with every session.
And this is a pretty fascinating development for game creators.
Here's how it works.
Instead of simply analyzing your source code and guessing what might go wrong, Autopilot launches the game and starts playing.
It reads your project's code, learns the mechanics, interacts with the environment, and tests how everything behaves.
And yes, it actually controls the game.
It can navigate levels, fight enemies, interact with objects, explore systems, and attempt complex objectives.
Here's what Autopilot can do:
Autonomous gameplay: Play your game through simulated keyboard and mouse input.
Bug detection: Explore levels and identify broken mechanics, unexpected behavior, and gameplay issues.
Feature verification: Test specific functionality and check whether it behaves as intended.
Goal-based testing: Complete objectives such as defeating a boss or finishing a level.
Code analysis: Study the project to understand how game systems work.
Persistent memory: Remember previous attempts and improve its strategies over time.
Testing reports: Document findings and report issues after each session.
And here's the really interesting part.
Autopilot doesn't approach every test as a completely new problem.
It learns from previous gameplay attempts.
The first time it encounters a game, it might struggle with navigation, combat, or unfamiliar mechanics.
But as it explores, it builds knowledge about the game and improves its decision-making.
The developers describe it as a continually learning playtesting system.
And the benchmark results are promising.
According to the developers' tests:
Autopilot: Completed 7 out of 7 game-testing tasks.
Claude Code: Completed 5 out of 7.
Another game-testing agent: Completed 4 out of 7.
That's a 100% success rate versus 71.4% for Claude Code in this small evaluation.
Autopilot was also approximately 1.7× faster than Claude Code, averaging 262 seconds per task instead of 437.
The team separately reports a threefold speed improvement over its previous Aura Verification Agent.
Of course, these are developer-reported benchmarks involving seven tasks in one Unreal Engine project, not proof of universal superiority.
Four ways to test your games:
1. Exploration
"Explore the crafting system and report anything unusual."
2. Testing
"Play through this level and identify all visible gameplay issues."
3. Verification
"Check whether the tree shakes when the player hits it with an axe."
4. Goal
"Defeat the boss without letting your health drop below 20%."
Autopilot launches the game, attempts the task, and generates a report.
And it works with the major game engines:
Unreal Engine 5
Unity 6
Godot 4
Windows and macOS
Want to use it with Claude Code or Codex?
After installing Autopilot and completing its setup, connect your preferred coding agent.
For Claude Code:
autopilot connect claude
For OpenAI Codex:
autopilot connect codex
Then simply ask: "Use Autopilot to explore my current game level, test its mechanics, and report any bugs you find."
The plugin is open source under Apache 2.0. It can use your existing Claude Code or Codex subscription for the coding-agent component, although the decision-model service has its own usage requirements. It's not an entirely offline local AI system. But here's what I find particularly exciting.
We've spent the last few years teaching AI agents to write games. They can generate characters, create environments, implement mechanics, and build entire playable prototypes.
Now they're starting to test those games themselves. Imagine asking an agent to implement a new combat mechanic.
Discovers that the player gets stuck during an attack animation. Then reports exactly what happened so the problem can be fixed. That's a much more interesting development workflow than simply asking AI to write code and hoping everything works.
Of course, automated testing won't replace experienced QA teams, human playtesters, or actual game design judgment. But it could make testing faster, more frequent, and much more accessible to independent developers. And there's something wonderfully absurd about the idea.
AI builds your game. AI plays your game. AI finds the bugs.
At some point, we might need to remind the agents that humans would occasionally like to play too.
Autopilot on GitHub: https://github.com/ramenvr/autopilot
Download and installation: https://www.tryaura.dev/autopilot/
Official documentation: https://www.tryaura.dev/documentation/autopilot
#Autopilot #AIGameTesting #ClaudeCode #OpenAICodex #UnrealEngine #Unity3D #Godot #AIAgents #GameDevelopment
YouTube / ABVX