Building production AI voice agents is a time-consuming process, largely due to wiring and testing, not prompt engineering. The complexity arises from integrating custom functions, calendar and CRM systems, and handling numerous edge cases. Manually testing these agents through countless call scenarios is inefficient and slow.To address this, a pipeline was developed using AI coding tools to automate these tasks. Claude Code generates the agent's structure and wiring from a simple specification. This includes defining custom functions and setting up the underlying workflow. The specification details the agent's purpose, capabilities, data collection needs, and desired tone.Subsequently, Comet, an AI browser automation tool, tests the generated agent. It simulates dozens of challenging call scenarios, mimicking real user interactions. These scenarios include interruptions, silence, off-script questions, and aggressive behavior. Comet analyzes transcripts and post-call data to identify where the agent fails.This automated loop replaces manual testing, allowing for rapid iteration. If an agent fails a test, the spec or flow is adjusted, and the relevant part is regenerated or edited. The pipeline significantly speeds up the process from initial concept to a robust, testable draft.However, human oversight remains crucial for critical decisions. Judgement regarding escalation boundaries, safety protocols, and compliance relies on human expertise. Automated analysis cannot fully capture nuances like robotic tone or an agent's responsiveness. Additionally, real-world processes like compliance registration and phone number provisioning are not affected by code generation.The pipeline's primary benefit is accelerating the non-core aspects of AI agent development. It frees up human time for high-value judgment calls that ensure trustworthiness. This automation explains why some AI voice builds are completed in days while others take months. The key differentiator is the automation of the development and testing loop.