Let Your Agent Cook: Using Skills to Evaluate and Improve Your App — Ankur Duggal, Arize AI

This post details how to use an AI coding agent, integrated with Arize's CLI and skills, to evaluate and improve an agentic application. It demonstrates connecting agent failures (e.g., invented answers, repeated tool calls, incorrect summaries) to specific traces and evaluation results. The agent is shown to investigate these traces, identify issues like grounding problems, fabrication, and silent failures, and then propose code changes. The workflow includes using dataset, evaluator, and experiment skills to make these proposed fixes testable, allowing for iterative improvement. The talk emphasizes the importance of tailoring evaluations to the application's specific failures and recommends human review of proposed changes to detect future regressions.