本文由 AI 分析生成
建立時間： 2026-09-05 來源： https://x.com/AndrewYNg/status/2095890279865721217
Summary
Andrew Ng breaks down “using coding agents” into five sub-skills that map onto the plan → execute → deploy workflow: directing the workflow, enabling agent autonomy, reviewing the work, customizing the agent’s environment, and understanding coding agent foundations. He argues the discipline has shifted from writing code to deciding what to build, calibrating autonomy, and verifying output.
Andrew Ng 將「使用編程 Agent」這項技能拆解成五個對應規劃→執行→部署工作流的子技能：引導工作流程、給予 Agent 自主權、審查產出、客製化 Agent 與環境、理解編程 Agent 的基礎原理。他認為這門學科的重心已經從寫程式轉移到決定要做什麼、校準自主程度、以及驗證輸出。
Key Points
- High-level workflow unchanged from pre-agent software dev (plan → execute → deploy/monitor), but effort shifts away from writing code toward spec-writing, architecture, and verification
- “Directing the workflow” = knowing how much human vs. agent effort each step needs, and when to loop back to an earlier step
- “Enabling agent autonomy” includes deciding interactive vs. delegated execution, managing context across a build’s phases, and running multiple agents in parallel with human/higher-level-agent orchestration
- “Reviewing the work” requires designing verification matched to the task — behavioral + functional checks, screenshots as evidence, eval sets with LLM-as-judge for qualitative work
- “Customizing the agent and environment” covers skills/plugins/MCP servers, hooks for automating repeatable steps, and maintaining living context files (AGENTS.md/CLAUDE.md)
- “Coding agent foundations” = understanding retrieval, context management, subagent interaction, and the harness-wrapped-around-an-LLM model, so you can recognize failure modes like overengineering or agents stopping short of the goal
- Warns against social media’s oversimplified narrative of long-horizon autonomous runs — most effective use is highly iterative with skilled human judgment
Insights
- Frames coding-agent skill as a five-layer competency stack rather than a single “prompting” skill — useful as a self-assessment rubric for where an individual or team is weak (e.g., strong at execution but weak at review/verification)
- The claim that practical utility of very-long-horizon autonomous runs is “amplified beyond reality” relative to cost is a useful counterweight to hype around agents running for hours unattended
- Pairs well with the “Stop Using /init for AGENTS.md” clipping analyzed in this same batch — Ng’s “customizing the agent and environment” skill and Osmani’s AGENTS.md research are directly complementary (what to maintain vs. how to write it well)
Connections
Raw Excerpt
Your skill at using coding agents will make you an effective builder. This positions you to also steer the overall build.