← Patterns

Vibe Coding to Agentic Engineering

Pattern · design-philosophy

Core Concept

“Vibe coding” — free-form prompt → code, with the human reviewing each output — is the toy phase of AI-assisted development.

“Agentic engineering” — structured workflows for AI agents across the full SDLC — is the mature form.

The shift moves the programmer’s job from writing and reviewing code to designing the workflow and defining “good enough.”

“A lot of time was spent discussing how to move from vibe coding to agentic engineering, by structuring workflows for AI agents across the SDLC. For example, building agent prompts for ‘ility requirements’ (reliability, flexibility, stability, scalability), running them in sequence for code review. The questions we’re all facing are ‘What should the workflow be?’ and ‘how do we know when it’s good enough?’” — Robbie McKinstry, DevGuide conference takeaway (May 2026). See source.

The Shift

Vibe CodingAgentic Engineering
Unit of workPrompt → codeWorkflow → artifact
Human roleAuthor + reviewerWorkflow designer + evaluator
Quality mechanism”Looks right to me”Defined acceptance criteria run as agent prompts
SDLC postureSingle-stageMulti-stage, composable
Failure modeOutput looks plausible but breaks in prodWorkflow gaps — unevaluated dimensions slip through
Skill being developedPrompt craftingSpecification, evaluation, workflow design

The concrete example from the source: instead of one prompt that writes code, you have a pipeline of agent prompts. One prompt checks reliability. Another checks flexibility. Another stability, another scalability. Each is a distinct evaluation stage. The human composes the pipeline and defines the pass criteria.

The Two Open Questions

Per the source, the entire industry is wrestling with:

  1. “What should the workflow be?” — There is no canonical SDLC-for-agents yet. Every shop is building its own. The room of Tailscale/Sentry/Honeycomb founders didn’t have a shared answer.

  2. “How do we know when it’s good enough?” — Code that “compiles and passes tests” used to be the floor. Now it’s the ceiling of vibe coding. The new floor needs explicit ility coverage, security review, performance budgets, operational fitness — each as an evaluation stage with a defined pass threshold.

Whoever answers these questions cleanly first will define the practice. This is why the SDLC-for-agents space is currently a land grab.

The Role Survival Paradox

The same DevGuide consensus held that code review by humans goes away, but programming as a profession does not. This is not a contradiction — it’s a role shift:

The job moves upstream from “author of code” to “architect of the production system that authors code.” See Bottleneck-Shifts-Upstream.

How to Apply

For any AI-assisted development effort, ask:

  1. What stages does this work need? (Discovery, design, test gen, implementation, review, QA, deploy.)
  2. What’s the agent prompt for each stage? (Including the implicit “ility” prompts: reliability, security, performance, accessibility, etc.)
  3. What’s the acceptance criterion at each stage? (Concrete pass/fail, not “looks good.”)
  4. Where are the human checkpoints, and what’s the human evaluating? (The artifact, not the code.)
  5. What’s reused vs. bespoke per project? (Workflows compound when they’re reusable.)

If you can’t answer 1–4, you’re still vibe coding regardless of which model you’re using.

Where I’ve Seen It

Cross-References