Loop Engineering: What "I Don't Prompt Anymore" Actually Means for Your Team
Boris Cherny says he writes loops, not prompts. What loop engineering really is, where it works today, and where the production claims outrun reality.
The head of Claude Code says he has stopped prompting AI and now writes loops that prompt it for him. It is a real shift, and it is where agentic development is going. It is also being described with more certainty than production reality supports.
From prompts to loops
Boris Cherny, who leads Claude Code at Anthropic, has described dropping one-off prompting in favour of loops — programs that run continuously, feed the model the current task, history, and goal, and let it plan, call tools, and act. Addy Osmani, who popularised the "loop engineering" label, breaks the pattern into five parts: scheduled automations that find and triage work, isolated worktrees for parallel runs, skills that encode project knowledge once, plugins and MCP connectors for tools, and sub-agents that separate the writer from the verifier. The difference from a prompt is simple: a prompt gives you one output; a loop gives you a compounding operation.
Why it isn't hype
On the right tasks this is genuinely effective: anything well-bounded, verifiable, and repetitive is a good candidate. Cherny's own examples — auto-rebasing, babysitting pull requests to production, periodic cleanups — are exactly that shape. The mental model is sound: stop being the thing inside the loop, and design the loop instead. In our own work, loops already handle pull-request hygiene, dependency bumps, and triage.
Where the framing outruns reality
Here is the part that gets compressed in the retelling. Cherny demos loops on his own repositories, with himself as the verifier, inside the company that builds the tool. The figures that travel with the story — engineers shipping "8x more code," Claude authoring "80%+ of merged production code" — are self-reported, on Anthropic's own codebases. They may be true there and still not transfer to a regulated enterprise with code-review gates, audit requirements, and a team that has to understand what shipped.
Osmani, notably, does not oversell it either. He lists the same caveats anyone running this at scale hits: verification stays manual ("unattended loops make unattended mistakes"), comprehension debt accumulates as you ship faster than you understand, and the real trap is cognitive surrender — letting the comfort of automation replace engineering judgment. In a production environment, those are not footnotes. They are the whole game.
And it is worth saying plainly, as opinion: the people most loudly declaring coding "basically solved" tend to have the most to gain from you believing it. Anthropic filed to go public on June 1. That does not make the workflow fake — it makes "read the confidence level carefully" good advice.
How it actually gets used
The pattern that works is narrow. Loops go on tasks where the cost of a wrong answer is low and the check is cheap: pull-request hygiene, dependency bumps, triage, test scaffolding. A human stays the verifier on anything that touches production logic. Evaluations come before autonomy, not after.
Loop engineering is real and it is directional — agentic development is moving this way. The mistake worth avoiding is reading a founder's demo as a production standard. Used narrowly, on bounded and verifiable work, it pays off today; handed the whole codebase, it just ships mistakes faster.
Sources: Addy Osmani (Loop Engineering); Digg and productmarketfit.tech (Cherny quotes); shanraisshan/claude-code-best-practice (June 2026).