
Preflight is an AI Agent QA and Reliability Harness that helps teams detect and prevent AI agent failures before they reach production. AI agents are becoming increasingly capable, performing complex tasks through tool usage, decision-making, and multi-step reasoning. However, they can fail in unpredictable ways through hallucinations, unsafe actions, incorrect tool calls, prompt injection attacks, and unexpected behavior across multi-turn conversations. Developers can connect their AI agent endpoint to Preflight, where the platform automatically generates diverse testing scenarios, including normal interactions, ambiguous requests, edge cases, and adversarial attacks. These scenarios are executed through simulated multi-turn conversations to evaluate how the agent behaves under realistic conditions. Preflight uses a dual AI judge pipeline to analyze each interaction. A fine-tuned Gemma 3 safety judge, trained on AMD infrastructure, provides domain-specific safety evaluation, while DeepSeek V4 Pro serves as an independent reasoning verifier. Together, the models evaluate task success, hallucination risks, instruction following, safety issues, and prompt injection vulnerabilities. The final verdict is determined through consensus logic: if both judges classify an interaction as unsafe, it is blocked; if both classify it as safe, it is allowed; and if they disagree, the case is flagged for human review. Preflight produces a reliability report containing an overall reliability score, failure statistics, and detailed failure transcripts. By leveraging AMD Developer Cloud and ROCm-powered infrastructure, Preflight enables scalable parallel testing and supports the fine-tuning and deployment of specialized AI evaluation models. Our vision is to provide developers and organizations with a pre-deployment reliability layer for AI agents, helping teams ship safer, more trustworthy AI systems with confidence.
13 Jul 2026