Woofun AI reports that Jensen Huang has declared the era of prompt engineering obsolete, asserting that "Prompt is dead; loop has taken its place" as the defining trend for artificial intelligence development. Huang emphasizes that the industry focus has shifted from manually crafting instructions to designing and managing autonomous loops where systems handle task initiation, execution, and verification without human input. This transition marks a fundamental change in how humans interact with AI, moving from direct command issuance to the role of rule designers who define goals and let the system operate until success or budget limits are reached. The concept of a "loop" refers to a self-sustaining cycle where the AI system manages the entire workflow, effectively replacing the human bottleneck identified in previous programming models. This new paradigm is gaining traction among industry leaders, including Peter, known as the "Father of Lobsters," Boris Cherny, the "Father of Claude Code," and Andrew Ng, who are all actively promoting Loop Engineering as the next standard for AI deployment.

The distinction between Loop Engineering and existing agent systems lies in the mechanism of continuous operation rather than the execution of individual tasks. While an agent performs the actual work, Loop Engineering provides the structural framework that allows the agent to function autonomously over extended periods without supervision. Without this loop mechanism, an agent remains a tool dependent on human instructions, but with it, the system evolves into a self-driving entity capable of independent problem-solving. Boris Cherny illustrates this shift by noting that he no longer writes prompts for Claude; instead, he maintains multiple concurrent loops that issue instructions and determine subsequent actions, reducing his role to simply writing the loops themselves. Similarly, Peter advises developers to stop writing prompts for programming agents and instead design loops that guide these agents through their workflows. This collective shift among experts indicates that Loop Engineering has transcended being a mere buzzword to become a practical necessity for scaling AI applications.

To understand the mechanics of this evolution, one must examine the previous standard process for AI programming over the last two years, which involved a repetitive cycle of human intervention. In that model, a human would write a prompt, the AI would generate code, the human would review and adjust it, and the process would repeat until satisfaction was achieved. Kapasi previously highlighted that humans were the bottleneck in this workflow and advised stepping back from direct involvement in every detail, a principle that Loop Engineering now fully embodies. The core logic of this new approach is summarized by the ability to define a goal and allow the AI to handle the rest, continuing its work until it succeeds or hits a predefined budget limit. This structural change redefines the human role from a 'messenger' relaying commands to a 'rule designer' setting the parameters for autonomous operation.

In practice, Loop Engineering is already embedded in familiar systems, with two dominant approaches currently leading the field: Claude Code and OpenAI Codex. Claude Code implements a three-component framework based on Loop Engineering principles, utilizing /loop for scheduled cycles, /goal to ensure tasks meet predefined criteria, and /schedule to manage automated cloud tasks. The most critical aspect of this design is the /goal component, which enforces the principle that a system should not evaluate its own performance; instead, Claude Code uses a large model to write code while an independent model, Haiku, checks it, ensuring objective evaluation. OpenAI Codex takes a different approach by combining automated pipelines, goal-driven mechanisms, and multiple sub-agents, with some developers reporting up to eight agents running simultaneously in cloud sandboxes to perform distinct tasks before merging results. Despite these architectural differences, both systems achieve similar outcomes by breaking complex tasks into smaller pieces, assigning them to multiple agents for parallel processing, and combining the results, suggesting that the real differentiation lies in the upper-level loop management strategies.

Boris Cherny's personal workflow serves as a concrete example of how Loop Engineering transforms daily operations in the software industry. He stated that he uninstalled his IDE last November and did not use it for a full month before deleting it completely, relying entirely on hundreds of sub-agents working simultaneously under his control. These agents perform diverse functions, including monitoring GitHub issues, analyzing user feedback on Slack, and checking for CI failures, with each agent operating in its own isolated code branch where one writes code and another runs tests and performs acceptance checks. Requests only reach his inbox when problems arise that require human judgment, and since the release of Opus 4.5, all his code has been written using Claude Code, with most of it now created directly on his mobile phone. This setup demonstrates the ultimate form of Loop Engineering where agents prompt each other without human intervention, allowing humans to simply define rules while the system handles the execution.

For developers looking to implement Loop Engineering, a blogger named Codez on X Corp has outlined a five-step process to ensure successful deployment. Step 1 requires conducting a 'four-question test' before building anything to avoid wasted effort, asking if the task occurs repeatedly, if there are automated acceptance mechanisms, if the token budget can handle the workload, and if the agent has the necessary tools. Only if all four answers are positive should the loop be built. Step 2 involves starting with the smallest possible loop structure, creating a basic three-component framework consisting of a Trigger (such as a scheduled task or event), a Skill (including project context in a file like STATE.md), and a State File (using Markdown to record progress and pending tasks). Step 3 mandates separating code-writing from evaluation tasks, using one model to generate code and an independent model to check it, ensuring the evaluator does not have access to the reasoning process of the coding model to prevent leniency. Step 4 focuses on avoiding common mistakes by setting hard stop conditions based on budget or iteration limits, saving state in files like STATE.md to preserve memory across runs, and restricting Loop Engineering to tasks with clear right-or-wrong outcomes like automatic code fixes or CI failure classification. Step 5 advises focusing on a single metric: the average cost per accepted change, noting that an acceptance rate below 50% indicates wasted resources as humans are still doing the review work the system was meant to eliminate.

The rapid rise of Loop Engineering follows a clear evolutionary path summarized by experts as Prompt, Context, Harness, and Loop, representing four distinct stages of AI development. From 2023 to 2024, prompt engineering dominated the field, where the quality of prompts determined AI performance and humans had to specify every detail. As model capabilities improved and context windows expanded, the industry shifted to 'Context Engineering' between 2024 and 2025, focusing on providing comprehensive information rather than just asking questions. By 2025 to 2026, the emergence of 'Harness Engineering' addressed the need for AI to interact with tools, execute code, and access APIs within a constrained environment. Loop Engineering now represents the latest evolution, focusing on whether AI can work continuously in this environment and advance tasks without human supervision, shifting the core requirement from executing a single task to maintaining an effective closed-loop system.

Although Loop Engineering has only recently gained industrial popularity, similar concepts have existed in academia for years, with significant contributions from Yao Shunyu of Tencent. His 2022 work on the ReAct framework, which won an Oral award at ICLR 2023 and has been cited thousands of times, combined 'reasoning' and 'action' into a continuous loop process where large models think, act, observe results, and repeat. This structure represents the earliest systematic expression of the 'agent loop' concept, which later evolved through mechanisms like Reflexion for error learning, Tree of Thoughts for multi-path search, and various tool-use agent studies that refined the 'planning + execution + feedback' cycle. These academic developments gradually converged into the 'loop systems' now recognized in the engineering community, demonstrating that Loop Engineering is not a singular invention but a gradual evolutionary path where a well-known Chinese expert played a pivotal role.

Despite the excitement surrounding this rapid development, concerns remain regarding the practical implications and costs of widespread adoption. Addy Osmani, the senior engineer at Google who named Loop Engineering, expressed caution in a detailed article, stating, 'It's still very early. I'm cautious. You need to be very careful with the token costs.' Kapasi offered an even more thought-provoking perspective at the Sequoia Capital AI Ascent 2026 conference, quoting a statement that has echoed in his mind: 'You can outsource your thinking, but you can't outsource your understanding.' This sentiment underscores the sober reality that while AI can help find solutions, humans must still truly understand the problems themselves. The shift from Prompt Engineering to Loop Engineering represents a continuous process of liberating humans from direct task management, yet it also introduces new challenges related to token costs and the potential loss of deep understanding within the codebase. This marks a critical juncture where the industry must balance the efficiency of autonomous systems with the necessity of human oversight and comprehension.