Vibe Coding and Agentic Engineering Are Converging, and That Raises Real Concerns

The distinction between vibe coding — generating code without deeply reading it — and agentic engineering — autonomous AI systems taking actions across tools and environments — is collapsing. Simon Willison flags this convergence as a problem worth taking seriously.

The concern is straightforward. Vibe coding trades verification for speed. Agentic systems trade human-in-the-loop control for autonomy. When these two modes combine, you get AI agents writing and executing code that no engineer has meaningfully reviewed. The blast radius of a mistake grows substantially.

For solo founders and small teams, the short-term productivity case is obvious. Agentic workflows that can scaffold, test, and deploy without constant intervention are appealing when headcount is limited. But Willison's framing is a useful corrective: moving fast on code you do not understand, inside systems that act autonomously, compounds the underlying risks of each approach rather than averaging them out.

The practical implication for engineers is about scope gating. Agentic systems benefit from well-defined boundaries — limited filesystem access, sandboxed execution, explicit confirmation steps before irreversible actions. When those constraints exist, autonomy is manageable. When they do not, vibe-coded agentic workflows become difficult to audit after the fact.

This is not an argument against either approach. Agentic tooling is improving fast, and the productivity gains at the task level are real. But the combination demands more discipline around system design, not less. Defining what an agent is allowed to touch, and what requires human sign-off, is now a first-class architectural decision.

Engineers adopting agentic workflows should treat scope definition the same way they treat API surface design: small, explicit, and versioned. The alternative is debugging production issues in code that nobody on the team wrote or reviewed.

Source