Andrew Ng: Using coding agents shifts attention from writing code to spec and verification

Andrew Ng's AI Engineering Skills Map series details the use of coding agents. The high-level workflow is Planning (brainstorming, spec writing, plan review), Execution (building, testing, verifying with calibrated autonomy), and Deployment/Monitoring (CI/CD, agent-monitored fixes). The attention shifts from writing code to deciding what to build, designing architecture, and validating outputs. Five key skills are: (1) Directing the workflow—deciding human vs. agent effort and when to iterate back. (2) Enabling agent autonomy—choosing autonomy level, managing context, parallelization, and safe operation. (3) Reviewing work—using tests, screenshots, LLM-as-judge, automated checks, and human review focused on code behavior rather than code. (4) Customizing agent and environment—maintaining AGENTS.md, hooks, skills, persistent state, and coordination across developers. (5) Coding agent foundations—understanding retrieval, context windows, and failure modes like over-engineering or stopping early. Ng criticizes social media narratives that overhype long autonomous runs; effective use requires high-judgment intervention at the right moments.