Ng: agents reshape engineering

Andrew Ng says artificial intelligence agents are changing software engineering fast enough that deciding what to build is becoming the harder job. (deeplearning.ai) Ng made that case in a June 17, 2025 talk at Y Combinator’s AI Startup School in San Francisco, where he said “execution speed” is a strong predictor of startup success and that AI practices are changing every “two to three months.” (ycombinator.com) He described “agentic” systems as software that can carry out multi-step work with limited supervision, and called the rise of agentic AI the most important technical trend he was watching in 2025. (singjupost.com) In Ng’s framing, coding is getting cheaper and faster, so the constraint moves upstream to product management: understanding users, choosing features, and making decisions quickly enough to match the new pace of building. (deeplearning.ai) He wrote on July 16, 2025 that product management is “the art and science of deciding what to build,” and said teams using agentic coders now value product managers with “very high user empathy” who can make fast calls. (deeplearning.ai) Ng has tied that shift to startup math. In an August 21, 2025 “No Priors” episode, he said work that once took six engineers three months can now be built “on a weekend,” turning a week-long wait for user feedback into a costly delay. (africa.businessinsider.com) That argument lands in a broader debate over what software teams should look like when code generation speeds up. Business Insider noted that some executives have cut or consolidated product roles to speed decisions, while Microsoft Chief Technology Officer Kevin Scott argued product managers still matter because they set the feedback loops that improve agents. (africa.businessinsider.com) Ng’s own public writing points in the opposite direction from “fewer product people.” He said some teams may need faster, more intuitive product judgment, because surveys, focus groups, and A/B tests help, but waiting for each one can slow a team that can already ship prototypes in a day. (deeplearning.ai) He has also been pushing training around those workflows. DeepLearning.AI now offers an “Agentic AI” course taught by Ng that teaches developers to build multi-step agent systems and includes testing, error analysis, and production optimization. (learn.deeplearning.ai) The open questions Ng is surfacing are less about whether agents can write code and more about who defines the work, who reviews it, and what schools should teach before graduates join teams that build with agents by default. His message has stayed consistent from mid-2025 into 2026: software is getting faster, but human judgment is not getting automated away. (ycombinator.com)