Agentic-Agile: Why building AI agents needs Agile, not just prompts

An article argues that prompt-driven workflows—opening a chat, writing a prompt, reviewing output, and re‑prompting—are a common but unreliable way to build AI agents. Drawing on experiments with M365 Copilot and GitHub Copilot CLI, it describes results from this approach as predictably non-deterministic and prone to emergent bugs no one specified or tested.