Andrew Ng on Coding Agents and Deciding What to Build

Andrew Ng discusses how coding agents change the skills developers need: less time implementing each detail, but more responsibility for engineering tradeoffs, customer feedback, and deciding what to build. His advice varies by project stage rather than prescribing one planning workflow for every task.

Based on the video's English auto-generated captions. Research and labor-market claims are Ng's account, not independently verified findings; his discussion of education products includes organizations he works with.

Key points

• AI engineering involves more than operating a coding agent. Ng describes four areas in DeepLearning.AI's skills map: building and deploying AI applications, software engineering fundamentals, using coding agents, and "shaping the build," meaning influencing both product decisions and the development process. He says the map draws on more than 10,000 job postings, structured interviews, surveys, and partner input. These are skills for several engineering roles, not only jobs titled AI engineer. 02:55–04:51(opens in a new tab)
• Human context can determine which decisions to retain. Ng expects agents to take over more standardized prompting and agent-management work. He contrasts that with decisions informed by context the agent lacks, such as a customer's reaction in a conversation or remembered market knowledge. His proposed dividing line is whether the human has relevant information unavailable to the agent, not whether a task currently requires elaborate prompts. 04:51–07:54(opens in a new tab)
• Uncertain outputs require systematic evaluation. Prompting a model or training a supervised system does not guarantee a known result in advance. Ng identifies monitoring, error analysis, and repeated evaluation as ways skilled developers make progress on these systems rather than treating each unexpected output as an isolated surprise. 07:54–09:58(opens in a new tab)
• Fundamentals help developers recognize hidden tradeoffs. Ng distinguishes rapidly changing coding-agent practices from slower-changing knowledge about deployment, monitoring, and engineering decisions. Without that knowledge, a developer may not notice that an agent has chosen between cost and latency or speed and reliability until the consequences appear later. He welcomes what inexperienced builders can now create but argues that engineering understanding still improves decisions. 12:51–15:38(opens in a new tab)
• Direct customer feedback can shorten the build loop. When implementation becomes faster, waiting for someone else to interpret user feedback can become a bottleneck. Ng argues that developers who can speak to customers and make some product decisions can keep iteration moving. He does not say engineers should become product managers: "shaping the build" also means deciding when to prototype, investigate a technology, seek feedback, or invest in architecture. 15:38–17:36(opens in a new tab)
• Match planning depth to project stage and reversibility. For early exploration, Ng favors quick prototypes that can be discarded, provided they do not expose sensitive data or create other unsafe outcomes. Mature work may need careful architecture, customer requirements, and security design. Long specifications and loosely directed overnight experiments both have a place; the mistake is applying either approach regardless of the project's risks and stage. 17:36–19:43(opens in a new tab)
• Completing work and learning are different objectives. Ng says studies of students show better homework performance but worse exam results when they offload thinking to language models, and he suspects working adults underestimate similar effects. He acknowledges that some uses of AI improve learning. His proposed response is to design experiences around developing the learner, rather than only producing an answer; he describes personalized education products as work in progress, not demonstrated outcomes. 19:43–23:53(opens in a new tab)

Full video: https://www.youtube.com/watch?v=qeMMnOiF51A(opens in a new tab)