This article, part of the “Exploring Gen AI” series on martinfowler.com, asks whether humans should stay out of the software development process and vibe code, or stay in it inspecting every line an agent writes. Its answer is a third position, which the author calls Humans on the loop: humans should build and manage the working loop rather than either leaving agents to it or micromanaging what they produce.
The argument rests on two loops. The why loop turns ideas into working software and iterates as ideas evolve; humans run it because they are the ones who want what it produces. The how loop is the process of building the software, in which intermediate artefacts — code, tests, tools, infrastructure, designs, ADRs — are created and used, and which in reality contains several nested loops, from specifying and delivering a feature down to generating and testing code. The article examines where humans sit relative to the how loop, and ends with a further step it calls the Agentic flywheel.
Key Points
- Intermediate artefacts, and the practices and patterns that produce them, are a means to the outcome people actually care about rather than deliverables in their own right.
- Humans outside the loop — leaving the how loop to agents — is described as the common definition of Vibe Coding, and the author says some interpretations of Spec-driven development are much the same.
- What matters is external quality (functional correctness and non-functional, operational quality such as not crashing, running fast, not leaking data, controlling hosting costs, passing compliance audits), and internal quality matters only when it affects external outcomes. The author argues it still does: agents understand and modify a cleanly designed, well-structured codebase more quickly and spiral less, and the time and cost of building systems matter.
- Humans in the loop, in the sense the article discusses, often means humans acting as gatekeepers inside the innermost loop, such as manually inspecting each line of code. The author argues this makes humans a bottleneck, since agents generate code faster than humans can inspect it, and suggests mixed reports on AI developer productivity may be partly due to people spending more time specifying and reviewing code than generation saves (see also Human-in-the-loop).
- Applying “shift left” thinking, agents produce better code when they can gauge its quality themselves rather than relying on humans to check it, so humans should instruct them on what is wanted and how best to achieve it.
- Humans on the loop means making agents better at producing results instead of inspecting what they produce. The collection of specifications, quality checks and workflow guidance that control the loops inside the how loop is, in the author’s terms, the agent’s harness, and building and maintaining it — Harness Engineering — is how humans work on the loop.
- The difference shows in what one does when unhappy with an agent’s output: in the loop, fix the artefact (directly or by telling the agent); on the loop, change the harness that produced it.
- In the agentic flywheel, humans direct agents to manage and improve the harness, starting from interactive consideration of their recommendations and possibly moving to automatic approval of recommendations with certain scores.
- The author suspects that for standard, frequently done work this may come to look like humans out of the loop, but argues that engineering the harness yields robust, “maybe even anti-fragile” systems that continuously improve themselves rather than one-off “good enough” solutions.
Context
The article presents the three positions as the author’s own framing. It notes that something like the on-the-loop concept has also been described as the “middle loop”, including by participants of a retreat on the future of software development, and it links to a separate martinfowler.com article on harness engineering for the practice itself. A footnote observes that the Ralph Loop is often used colloquially to mean leaving agents looping until they finish, whereas as originally described the operator plays an important role in steering them.