In 2026, the field of AI-assisted programming is undergoing a profound paradigm shift. The way developers interact with AI is evolving from the manual, turn-by-turn process of ‘Prompt Engineering’ to a more advanced, automated paradigm: ‘Loop Engineering.’ This transition, driven by several industry leaders, signals the future of software development.
From Human-in-the-Loop to Automated Systems
In the traditional model of AI programming collaboration, the developer acts as a dispatcher and quality inspector, guiding the AI by repeatedly inputting prompts and reviewing outputs—the human themselves forms the ‘loop.’ In November 2025, after relying entirely on AI to submit 259 pull requests in a single month, Boris Cherny, creator of Anthropic’s Claude Code, uninstalled his own Integrated Development Environment (IDE). On June 2, 2026, he further noted that he no longer writes prompts for Claude, but instead focuses on ‘writing loops,’ letting automated systems prompt the AI and iterate on tasks.
This idea quickly gained traction. OpenAI’s Peter Steinberger publicly urged developers to stop writing prompts and instead design ‘loops that prompt agents.’ Subsequently, Google’s senior engineering lead, Addy Osmani, formally named this concept ‘Loop Engineering,’ defining it as a higher-level abstraction above ‘Harness Engineering.’ The core idea is that developers are no longer the direct operators manipulating the AI, but rather the engineers who design and maintain the entire automated ‘production line.’
Core Components and Practices of Loop Engineering
Loop Engineering doesn’t completely abandon prompts; instead, it encapsulates them within reusable, automated workflows. According to Addy Osmani’s summary, a complete loop system consists of six core components:
- Automations: Triggered by timers or events, these proactively discover tasks (e.g., scanning an issue list) and assign them to AI agents.
- Isolated Workspaces (Worktrees): Using mechanisms similar to
git worktree, they provide separate code checkout environments for each concurrent AI task to avoid conflicts.
- Skills: Project specifications, build steps, and common problem solutions are documented (e.g., in a SKILL.md file) to provide stable context for the AI, helping it overcome its ‘forgetfulness’ problem.
- Connectors: These allow AI agents to interact with external systems (like databases, APIs, and ticketing systems), enabling them to execute complete workflows from code modification to project status updates.
- Sub-agents: Adopting the ‘separation of concerns’ principle, an independent review agent is used to evaluate and correct the work of another coding agent, thereby improving code quality.
- Memory: Persistent storage (such as a Markdown file) is used to record task progress, decisions, and outcomes, ensuring continuity for long-running tasks.
In practice, Anthropic’s Claude Code implements loop functionality through the /loop (timed repetition) and /goal (condition-driven) commands. A three-tiered system configuration called ‘THE HIVE’ demonstrates its potential: local loops handle immediate tasks, cloud-based Routines execute long-running background tasks, and a Cluster processes large-scale tasks in parallel via the /batch command.
Industry Convergence and Challenges
‘Loop Engineering’ is not an entirely new invention. Its ideas can be traced back to early agent projects like the ReAct framework and AutoGPT in 2022, as well as researcher Simon Willison’s ‘agent-in-a-loop’ design proposed in September 2025. The new trend in 2026, however, is that major AI programming tools are beginning to productize and build these components in. The high degree of similarity in related features between Anthropic’s Claude Code and OpenAI’s Codex indicates that this has become an industry consensus and a new competitive focal point.
Despite its promising outlook, Loop Engineering still faces three major challenges:
- Cost: Complex loop systems, especially those involving multiple agents and dynamic workflows, can lead to high token consumption.
- Runaway Risk: An autonomously running AI may exhibit problems like ‘shirking’ (not fully completing tasks), ‘overconfidence’ (overestimating the quality of its work), or ‘drift’ (deviating from initial instructions after many iterations). This highlights the importance of separating the ‘executor’ from the ‘inspector.’
- Conceptual Skepticism: Some developers believe ‘Loop Engineering’ is merely a rebranding of existing concepts and are skeptical of the actual value it brings.
Redefining the Developer’s Role: From Implementer to System Designer
The rise of Loop Engineering requires developers to shift from being direct creators of code to becoming managers and system designers of automated digital teams. Claire Vo, founder of ChatPRD, likens it to ‘onboarding a new employee,’ where the developer needs to define job objectives, processes, and acceptance criteria.
In this new paradigm, differences in model performance are gradually leveling out. The real ‘moat’ is no longer a cleverly crafted prompt, but the automated loop system a developer designs—one that can continuously iterate and accumulate knowledge. However, this also introduces a new risk: developers must be wary of ‘Comprehension Debt,’ the loss of deep understanding of the code they review and merge due to over-reliance on automation. Loops are a powerful lever, but only if the user clearly understands how they work and the quality of their output.
Ultimately, Boris Cherny’s confidence in uninstalling his IDE didn’t come from ‘not writing code’ or ‘not writing prompts,’ but from having solidified his own knowledge and workflow into a more powerful automated system. For all developers, embracing Loop Engineering means starting to build their own systems while always retaining final oversight and understanding of the system’s behavior.
