Forget prompt engineering: 'Loop engineering' is all the rage now

In the evolving landscape of AI, a fresh concept is gaining traction: loop engineering. Notably, Boris Cherny, the creator of Claude Code, recently revealed his shift away from traditional prompt writing. Instead, he relies on loops, stating, "It's an agent that prompts Claude." This innovation allows him to engage with an enhanced version of Claude that takes the lead in coordinating tasks. Cherny isn't alone in this movement. OpenAI engineer Peter Steinberger, known for his popular OpenClaw project, recently urged users to transition from manual prompts to designing loops for AI agents. He emphasized that the time for prompting coding agents is fading, suggesting that users should focus on crafting loops that effectively guide their agents. Loops serve as recurring structures that streamline user interactions with AI, minimizing the need for constant prompt input. For instance, a simple command like /goal allows AI tools, such as Anthropic's Claude Code or OpenAI's Codex, to autonomously continue working until a task is completed, reducing the user's burden. Claire Vo, founder of ChatPRD and host of the "How I AI" podcast, encapsulated the essence of this shift by stating that users no longer need to rely on manual input for their AI agents to function effectively. The consensus among experts like Addy Osmani, Google Cloud's director, is that the traditional method of directly prompting generative AI tools may soon become obsolete. Osmani outlined that effective loops should consist of five key elements: automations, worktrees, skills, plugins, and connectors, with automation being the cornerstone for ensuring the loop's repeatability. A common strategy in coding involves using one agent to write code while another reviews it, as Osmani noted, "The model that wrote the code is way too nice grading its own homework." Steinberger shared a practical example of a loop he utilizes, which involves instructing Codex to manage repositories and refresh every five minutes to direct the workflow. He highlighted how this method enables task parallelization and efficient management. While current discussions on loops primarily focus on coding, Vo pointed out that this concept has broader implications, suggesting that managers can view their AI agents as new employees, whether they are executive assistants or customer service representatives. In fact, many users may already be employing loops without realizing it, such as when utilizing scheduled tasks in ClaudeCowork. However, concerns about the cost of operating multiple AI agents persist, as running advanced AI models can quickly deplete personal token budgets. In response to inquiries about budget-friendly adjustments, Steinberger advised opting for less frequent API calls to conserve tokens while still maintaining functionality. He humorously noted his position as someone with access to abundant tokens, a perk of working at OpenAI. Ultimately, the key takeaway from this emerging trend is to allocate resources wisely, as Osmani cautioned that while subagents can enhance effectiveness, they also incur additional costs. As loop engineering continues to gain momentum, it stands to reshape how users interact with AI, making the process more efficient and less reliant on manual input.