Only believe what you can validate: a verification framework for agentic AI (devblogs.microsoft.com)

A new verification framework for agentic AI has been proposed to address the challenges of validating AI-generated outputs in critical applications, particularly those involved in application modernization. The framework is grounded in the premise that "only believe what you can validate," highlighting the common pitfalls of relying on seemingly coherent but potentially erroneous outputs generated by agentic AI systems. The traditional approach of manual inspection is impractical given the scale and complexity of AI outputs, which can often misinterpret or hallucinate based on the vast data they process. The iterative verification framework emphasizes collaboration among humans, AI, and deterministic tools to ensure the accuracy and reliability of agentic AI outputs. It outlines a structured approach consisting of specific checks, costs, performing agents, and outcomes to systematically assess the produced content. This framework not only seeks to improve verification efficacy but also aims to reduce expert fatigue and enhance confidence in AI applications where precision is paramount, thus paving the way for safer AI adoption in industries reliant on critical operational data.