This post starts from Vibe Coding, recounting the term’s coinage in February 2025 as a style of coding that surrenders to the “vibes” and forgets the code exists, and asserts that by 2026 it has become ordinary — to the point that services have appeared through which non-engineers publish more than thirty web tools built this way and the value of simply being able to write code is falling. Its question is what professional engineers should do instead, and its answer is Agentic Engineering.
The author defines agentic engineering as a style of development in which a person directs and coordinates multiple AI agents while retaining final responsibility for architecture and quality — in the post’s phrase, becoming the tech lead of an AI team rather than handing everything to the AI. The post contrasts the two approaches, illustrates a division of work into implementation, testing and refactoring agents under a human lead, and sets out five principles.
Key Points
- The post identifies three limits of vibe coding: prototypes that run but are not fit for production (missing error handling, security holes such as SQL injection and XSS, performance problems, unmaintainable code); generated code that is a black box nobody can later fix, so technical debt accumulates until a rewrite looks cheaper; and an inability to handle complex requirements such as integrating several systems, legacy code, bespoke business logic and scalability.
- Its comparison table casts the human in vibe coding as a requester and in agentic engineering as designer and supervisor; one-to-one dialogue versus orchestration of multiple agents; quality left to the AI versus human review and approval; prototypes versus production products; and prompting skill versus design, prompting and review skill together.
- The author names Claude Code as the most practical environment for this style as of February 2026.
- Principle 1: decompose work into appropriately small tasks, so that each step’s output can be checked and corrected, which the post calls the key to quality.
- Principle 2: pass context explicitly, structuring project background, constraints and conventions in files such as CLAUDE.md or a README.
- Principle 3: review is the human’s job; AI-generated code must not be merged as-is, with a checklist covering security holes, N+1 queries, edge cases, consistency with existing code and test sufficiency.
- Principle 4: design on the assumption that agents fail — feature branches, automated tests in CI, preview environments and staged rollouts — to detect failures early and limit their impact.
- Principle 5: keep developing one’s own expertise, since judging AI output depends on knowledge of architecture, security, performance and the domain; the post argues paradoxically that the better one uses AI, the more human expertise matters.
- The author concludes that an engineer’s value is moving from writing code quickly and correctly to guaranteeing the correctness of code an AI has written, and frames this as an evolution, comparing it to doctors using surgical robots and pilots using autopilot.
Context
The post is a short, polemical argument by an individual developer rather than a study, and its definition of agentic engineering is presented as the author’s own formulation. It closes with a short list of reference links.