Workflow Reframing: "Vibing" And Architecture-Heavy Decision Making

Issue 87 Edition 2026-03-28 5 min read

General

Sources: 1 • Confidence: High • Updated: 2026-03-29 03:24

Key takeaways

Matt Webb describes his current practice as "vibing" rather than "coding" or "vibe coding".
Agentic coding tends to solve problems by exhaustively iterating until the problem is eliminated, even at extremely high token and compute cost.
The desired outcome for AI coding agents is fast solutions that remain maintainable, adaptive, and composable so improvements elsewhere can lift the whole stack.
A strong foundation for agentic and developer productivity is high-quality libraries that encapsulate hard problems behind interfaces that make the correct approach the easiest approach.
In a "vibing" workflow, developers may read fewer lines of code while making more architecture-level decisions.

Matt Webb describes his current practice as "vibing" rather than "coding" or "vibe coding".
In a "vibing" workflow, developers may read fewer lines of code while making more architecture-level decisions.

Agentic coding tends to solve problems by exhaustively iterating until the problem is eliminated, even at extremely high token and compute cost.

The desired outcome for AI coding agents is fast solutions that remain maintainable, adaptive, and composable so improvements elsewhere can lift the whole stack.

A strong foundation for agentic and developer productivity is high-quality libraries that encapsulate hard problems behind interfaces that make the correct approach the easiest approach.

What is the empirical distribution of token usage, loop length, and run time for agentic coding sessions across representative tasks?
What is the marginal utility curve of additional agent iterations versus quality/maintainability outcomes (e.g., defects, rework, long-term change cost)?
Do maintainability, adaptability, and composability measurably improve or degrade under agentic coding compared to baseline development for the same systems?
How much of agentic productivity (and safety) variance is explained by the presence of high-quality shared libraries/interfaces versus ad-hoc generated implementations?
Is "vibing" an idiosyncratic label for one practitioner or an emerging, widely adopted workflow category with stable practices and responsibilities?

Rising need for cost controls in agentic coding as iteration can drive extreme token and compute usage, creating demand for monitoring, budgeting, and loop containment features in developer platforms.
Greater emphasis on lifecycle quality metrics for AI-generated code, as success is framed as maintainable, adaptive, composable outcomes, implying growth for tools that measure and enforce these properties beyond task completion.
Increased value of high-quality shared libraries and interfaces that constrain agents toward correct implementations, implying investment focus on ecosystems that provide robust primitives and make the right approach the easiest approach.

Published benchmarks showing token usage, loop length, and runtime distributions for agentic coding across representative tasks, with clear identification of runaway-loop frequency and cost drivers.
Evidence that maintainability, adaptability, and composability improve under agentic workflows, measured via defects, rework rates, and long-term change cost compared to baseline development on the same systems.
Demonstrations that projects with strong shared libraries and interfaces show lower variance in agent output quality and safety, and require fewer iterations to reach acceptable results.

Data showing agentic iteration costs are typically low and loops are short for most production tasks, reducing the importance of specialized cost and loop management.
Empirical results indicating agentic coding degrades maintainability or increases rework compared to baseline development, undermining the stated quality bar as achievable in practice.
Findings that high-quality shared libraries and interfaces do not materially affect agent productivity or output quality, implying benefits come mainly from ad-hoc generated implementations rather than reusable primitives.