Vibe Coding and Agentic Engineering Are Converging Faster Than Expected

The distinction between vibe coding and agentic engineering is getting harder to defend. Vibe coding, where a developer describes intent and accepts AI output without deep review, was positioned as a casual or beginner practice. Agentic engineering, where AI systems execute multi-step tasks with tool use and autonomy, was considered a serious engineering discipline. Willison's analysis is that these two modes are converging, and not in a direction that engineers should be comfortable with.

The concern is structural. As agentic systems become easier to invoke and models become more capable, the temptation to treat agentic pipelines the way vibe coders treat generated snippets grows. You describe a goal, the agent executes across files and services, and the developer reviews the diff — if at all. The feedback loop that made agentic engineering defensible starts to disappear.

For engineers building on top of LLM APIs or agent frameworks, this convergence matters operationally. Code that runs autonomously across a codebase or against live APIs carries different risk than a single completion you paste into a file. When the mental model shifts from review-then-accept to accept-then-notice-problems, the blast radius of a bad generation expands significantly.

For solo founders and small teams using agentic tools to move fast, the practical implication is that process discipline has to come from the builder, not the tooling. Current agent frameworks do not enforce review gates. The defaults favor throughput.

Willison's position is not that agentic engineering is wrong or that vibe coding has no place. The argument is about awareness: knowing which mode you are operating in and what assumptions each carries. When the two blur, engineers often inherit the risk profile of vibe coding while believing they are practicing something more rigorous.

The gap between confidence and control is where things break.

Source