AWS Unveils System for Continuous Evaluation of AI Agent "Kiro" Prompts Using Real Conversation Data and LLMs

Because an AI agent's system prompt operates within every possible combination of foundation models, tools, codebases, tasks, and users, it is difficult to exhaustively verify all behaviors using a test suite. Furthermore, as improvements in LLM performance lead to more stringent interpretation of instructions, "prompt staleness"—where prompts tuned for older models inadvertently cause negative impacts—has become a challenge.

The evaluation framework introduced by AWS is centered around an "LLM Judge" that scores internal conversation sessions across 15 aspects of behavioral quality. The judge makes determinations based on specific evidence, such as explicit user dissatisfaction, conversation abandonment, or task interruption; it is designed to exclude ambiguous conversations from the evaluation to avoid excessive false positives.

The company uses this mechanism to compare changes to system prompts through A/B testing. Their operations include screening 27 potential changes to exclude those with adverse effects. Experimental results showed significant improvements: explicit dissatisfaction signals decreased by 5% in the Kiro CLI, and behavioral quality issues decreased by 20% in the Kiro IDE.