Building an AI game testing agent with Amazon Bedrock

This article describes building an autonomous QA agent on Amazon Bedrock that connects to running Unity games and executes test scenarios through AI reasoning rather than brittle scripts or lengthy training.

• LLM-based agent reads plain-language test cases, reasons about game state, and decides next actions without fragile scripts or reinforcement learning overhead
• Architecture uses Amazon ECS/Fargate for the agent, AltTester® for game instrumentation, DynamoDB for test data, and Amazon Bedrock Knowledge Bases for game-specific context
• TITAN reasoning loop implements perceive-reason-act-reflect cycle: perception filters 1400+ objects to 42 meaningful ones, Claude selects from 13 game-agnostic tools, reflection detects stuck patterns
• Game-agnostic tools like tap, read_text, find_highlighted_tiles separate game knowledge from execution logic through parameterized design
• Web UI powered by CloudFront, Cognito, and AppSync Events lets testers submit tests and watch live execution traces with cost breakdowns
• Prototype validated on turn-based mobile strategy game; 50 nightly tests cost ~$760/month, scaling to $3 per test at 1000 tests/night

The approach enables autonomous QA that scales with game complexity without rearchitecting, freeing testers for exploratory testing while catching bugs earlier in release cycles.