Summary
This first-place entry in DeepLearningAI’s 7-day Voice AI Builder Challenge demonstrates a coding agent that autonomously handles routine code fixes and escalates genuine judgment-call decisions to the developer via an outbound phone call. Built by Nikolaos from Greece using Vocal Bridge and a Claude-powered CLI running fully unattended, the demo targets a real checkout API with a validation bug duplicated across five handlers including create_order, apply_coupon, refund, and capture.
The agent independently resolves minor issues — tab/space inconsistencies, typos, and JSON defaultable choices — before identifying a decision it cannot make alone: whether to unify the validator into a clean refactor that breaks the existing error contract, or maintain backward compatibility with some duplication. Unable to determine which downstream clients depend on the current error format, the agent places a live phone call to the developer.
The resulting conversation showcases natural human-AI dialogue, with the AI pressing for a binary A/B choice, handling the developer’s hesitation, and ultimately locking in option B (clean refactor) after explicit confirmation. The agent then applies the change, logs the decision with context for future reference, and updates the Vocal Bridge dashboard — completing the entire workflow without the developer touching a keyboard.
📺 Source: DeepLearningAI · Published August 05, 2026
🏷️ Format: Showcase
