Could AI Agent Loops Make Prompting Obsolete?

The rise of generative artificial intelligence transformed how people interact with technology. Following the launch of ChatGPT in 2022, users largely relied on prompts to guide AI models. As a result, prompt engineering became a valuable skill, with carefully crafted instructions often producing better results.

However, the next phase of AI interaction is already taking shape. Developers are increasingly turning to AI agents that can complete tasks with limited human involvement. Instead of repeatedly entering prompts, users can assign work, review progress, and intervene only when necessary.

Now, another shift is emerging. Developers are creating automated systems that continuously guide AI agents until a task is finished. This approach, known as loop engineering, aims to reduce the need for direct human prompting and could significantly change how people work with AI systems.

The Shift From Prompting to Loops

Traditional AI workflows depend on users providing instructions at each stage. In contrast, loop engineering allows developers to build recurring systems that automatically generate the prompts required for AI agents to continue working.

Several AI leaders have highlighted the potential of this approach. Boris Cherny, head of Claude Code, said he no longer writes prompts himself. Instead, an AI agent creates prompts for Claude and coordinates the overall workflow.

Similarly, Peter Steinberger, creator of the OpenClaw project and an OpenAI engineer, has encouraged developers to stop manually prompting coding agents and instead design loops that handle prompting automatically.

Addy Osmani, director at Google Cloud, also argued that direct prompting of AI coding tools is becoming less important. According to him, loop engineering replaces the human as the person issuing prompts by creating a system that manages the interaction instead.

What Are AI Agent Loops?

An AI agent loop is a recurring system designed to guide agents through tasks without requiring continuous human input. Rather than manually directing the AI, users create a framework that repeatedly nudges agents towards completing a goal.

According to Osmani, loops typically include five main components:

• Automations, which allow tasks to run repeatedly instead of as one-off actions.
• Worktrees, which enable multiple AI agents to operate in parallel without duplicating effort.
• Skills, which provide instructions and project-specific knowledge.
• Plugins and connectors, which give agents access to existing tools and workflows.
• Sub-agents, which divide responsibilities between agents, such as creating work and reviewing it.

In addition, memory plays a crucial role. Developers often store progress and task information externally because AI models do not retain knowledge between separate runs.

How Developers Are Using Loops

Developers are already implementing loops in several ways. One example is a system that instructs AI coding tools to continue working until a specific objective is completed.

Steinberger described a Codex-based loop that maintains repositories, periodically checks for work, and distributes tasks across different threads. Consequently, teams can parallelise tasks while maintaining oversight.

Experts also recommend separating responsibilities among agents. For example, one agent may write code while another reviews the final output. This approach helps avoid overly favourable self-assessments from the same model that generated the work.

Moreover, loop engineering is not limited to software development. Claire Vo, host of the How I AI podcast, suggested that managers could create loops for recurring workplace processes, including employee onboarding and weekly productivity reviews.

Challenges and Limitations

Despite the potential benefits, loop engineering comes with significant challenges. Long-running loops often consume large numbers of tokens, increasing operational costs. Since multiple agents and sub-agents may work simultaneously, expenses can rise quickly.

To reduce costs, developers can limit how frequently loops run. For instance, tasks scheduled hourly or daily generally consume fewer resources than continuously running systems.

Osmani also advised using sub-agents selectively because each additional agent requires its own model and tool interactions. Therefore, organisations should reserve extra layers of review for situations where a second opinion delivers clear value.

Some users may also prefer scheduled workflows rather than continuous loops. Scheduled tasks run only at specific times, making them more suitable for cost-sensitive operations.

Security and oversight remain additional concerns. While automation can improve efficiency, experts caution against removing humans entirely from the process. Effective loop engineering still requires human judgement, supervision, and accountability to ensure reliable outcomes.

With inputs from Reuters