The Era of Loop Engineering: Why Boris Cherny Stopped Writing Prompts

Since the boom of Large Language Models (LLMs), prompt engineering has been widely hailed as the ultimate skill for developers communicating with AI. We became accustomed to meticulously crafting long system prompts and tweaking adjectives to prevent models from going off the rails.

However, Boris Cherny, the creator and lead of the Claude Code project at Anthropic, recently challenged this paradigm with a provocative statement: "I don't write prompts anymore. I write loops."

Welcome to the era of loop engineering. This is not merely a minor tool adjustment; it represents a fundamental shift in how we build and interact with software.

From Prompting to Loop Engineering

With traditional AI autocomplete assistants like GitHub Copilot, your workflow is often a series of repetitive manual steps:

1. You write a prompt or start coding.
2. The AI generates a snippet.
3. You manually copy-paste it into your IDE.
4. You run it, only to encounter a compiler or runtime error.
5. You write another prompt explaining the error and asking for a fix.

This loop positions the human as a manual router between the execution environment (terminal/compiler) and the AI.

Loop engineering eliminates this friction by automating the entire feedback loop. Instead of expecting constant human intervention, developers design autonomous agentic workflows. In this model, the AI agent reads directories, runs shell commands, triggers test suites, catches compiler errors, and refactors its own code until all pre-defined constraints are satisfied.

In other words, humans stop micro-managing the AI's inputs and outputs. Instead, we design the feedback loops within which the AI autonomously iterates toward a correct solution.

How Boris Cherny "Solved" Coding

Boris Cherny shared a striking statistic: Since November 2025, he has completely stopped manually typing code for his day-to-day programming tasks. Instead, he delegates everything to Claude Code — an agentic coding system that operates directly within his CLI.

By utilizing autonomous, closed-loop execution, Cherny routinely creates and merges 10 to 30 pull requests (PRs) per day. Anthropic also reported a massive increase in overall engineer productivity since adopting this agent-first workflow internally.

Metric	Traditional Prompting	Loop Engineering
Human Role	Writing granular prompts, copy-pasting, manual debugging	Defining high-level goals, sandboxing, reviewing PRs
Interaction Frequency	Constant (every few seconds or minutes)	Ephemeral (start and end of the task)
Self-Correction	Relies entirely on the developer's manual feedback	AI runs test suites, parses error logs, and auto-refactors
Output Velocity	Capped by human typing and reading speeds	Dozens of clean, tested PRs automatically delivered daily

The Impact on Developer Experience

For the software engineering community, this evolution brings both massive opportunities and a necessary shift in mindset.

1. Rising up the Abstraction Ladder

Previously, engineers took pride in mastering complex syntax, manual algorithm optimization, or configuring massive YAML files. With loop engineering, your job moves entirely to system architecture. You must clearly define: What security compliance standards must this application meet? What are the core business logic boundaries? How do I write a test suite robust enough to prevent the AI from taking lazy shortcuts?

2. The Absolute Reign of Automated Testing

An AI agent can only self-correct if it has a reliable "compass" to distinguish right from wrong. That compass is your test suite. If a project lacks unit or integration tests, the AI agent will inevitably introduce regressions or write hallucinated logic without realizing it. Writing tests is no longer a chore left for the end of a sprint; it is the fundamental enabler of AI-driven development.

3. A More Mature "Vibe Coding"

Many fear that "vibe coding" (building software purely with natural language) will lead to an unmaintainable mountain of technical debt. However, when paired with automated cleanup and refactoring loops, vibe coding actually eliminates repetitive boilerplate work, freeing up human engineers to solve truly complex system design problems.

Conclusion

While Cherny's claim that "coding is solved" might carry a bit of marketing hype to promote Anthropic's new agentic tools, the underlying trend is undeniable. Software developers of the near future will transition from being direct typists of code to becoming "loop designers" who orchestrate powerful AI agents.

If you are still spending hours tweaking long system prompts, it might be time to step back, look at the bigger picture, and ask yourself: How can I build a better feedback loop for my AI?

---

Content assisted by AI (Amy 🌸). Reviewed by the author.