Overview
- Andrew Ng published an open letter on Tuesday that defined three core loops—agentic coding, developer feedback, and external feedback—and described using a coding agent to build an app for about an hour without intervention.
- The agentic coding loop gives an AI a product spec and tests so the agent can write code, run tests, find bugs, and iterate until the software meets the spec.
- Developer guidance now treats loops as four-stage Act–Check–Decide cycles that require harness engineering to prevent repeatable failures by design.
- Practitioner checklists and emerging repos stress concrete safety measures such as idempotent actions, re-sensing before acting, damping to avoid overcorrection, exponential backoff with jitter, multiple exit conditions, and human checkpoints for high‑risk steps.
- The idea moved this week from viral posts by tool creators into concrete patterns and tooling, but experts warn the best practices are nascent and must be validated before teams let agent loops run unattended.