A useful name for a real workflow shift

On March 16, 2026, Hacker News pushed Simon Willison’s new guide chapter What is agentic engineering? to 151 points and 87 comments at crawl time. Willison uses the term agentic engineering to describe software development carried out with coding agents, meaning systems that can both generate code and execute it. His core definition is compact but useful: agents run tools in a loop to achieve a goal. For coding agents, the decisive tool is code execution.

That distinction matters because it separates ordinary chat-style code generation from systems that can actually test, inspect, and iterate. In the guide, Willison names tools such as Claude Code, OpenAI Codex, and Gemini CLI as examples of the new class of development interfaces. The point is not that the model writes perfect code unaided, but that it can repeatedly act on the environment and produce evidence about whether a change worked.

What humans still own

Willison’s argument is not that software engineering is shrinking to prompt-writing. He says the hard part remains deciding what code to build, which tools the agent should have, how much detail to provide, and how to verify whether the result is credible. He also emphasizes that language models do not automatically learn from previous failures, but agent workflows can improve when humans refine instructions, tools, and harnesses after each mistake.

That framing is more rigorous than the common “AI writes the app now” storyline. It treats agents as fast iterators inside a process that still needs specification, review, and quality control. The guide itself reflects that view: it is structured as a living document with sections on testing, manual QA, code walkthroughs, and annotated prompts, not as a claim that autonomous generation has replaced engineering discipline.

Why the distinction with vibe coding matters

Willison also draws a clear line between agentic engineering and vibe coding. He references Andrej Karpathy’s earlier phrase as useful for describing prototype-grade code where the author effectively “forgets that the code exists,” but argues that the term should not cover every serious use of coding models. In his version, production-intended work still requires review, iteration, and accountability. That is likely why the HN thread resonated: the conversation around coding agents is moving away from novelty demos and toward repeatable operating patterns.

The practical takeaway is that agent use is becoming an engineering method, not just a user interface trend. Naming that method helps teams talk about tooling, testing, and failure modes in concrete terms instead of collapsing everything into hype or backlash.

Primary source: Simon Willison guide. Community discussion: Hacker News.

#simon-willison

Hacker News Pushes “Agentic Engineering” Forward as Simon Willison Defines the Coding-Agent Workflow