Anthropic’s Boris Cherny, who leads Claude Code, and OpenClaw creator Peter Steinberger are independently making the same argument: developers should stop writing prompts and start building systems where AI agents write prompts for themselves. The convergence signals that “loop engineering,” where agents autonomously generate, refine, and execute prompts toward human-defined goals, is becoming the dominant architecture for production agent systems.
What Loop Engineering Means
A loop is a self-running AI workflow. Instead of a developer crafting individual prompts and iterating manually, the developer defines an objective, and the agent keeps working: checking its output, generating new prompts, refining results, and repeating until the task is complete.
“It’s an agent that prompts Claude. I don’t write the prompt anymore. Claude writes the prompt,” Cherny told Business Insider, as reported by The Economic Times.
Steinberger reinforced the point from the tooling side. “You shouldn’t be prompting coding agents anymore. You should be designing loops that prompt your agents,” he said, according to The Economic Times. OpenClaw’s /goal command is cited as a concrete example: users define an objective, and the agent continuously works toward it while self-managing its own prompting logic.
Google Cloud executive Addy Osmani described effective loops as systems that combine automation, plugins, connectors, worktrees, skills, and specialized sub-agents working together, according to The Indian Express.
From Tool Use to Team Management
ChatPRD founder Claire Vo framed the shift in organizational terms. “This is the time for the manager. You are designing a job,” she told The Economic Times. In loop-based systems, one agent might write code, another reviews it for bugs, and a third manages documentation or deployment. The human’s role shifts from writing instructions to designing workflows and supervising execution.
This reframes the developer’s relationship with AI. Prompt engineering treated models as tools that needed precise instructions. Loop engineering treats them as workers that need job descriptions, feedback mechanisms, and clear success criteria.
The Cost Question
Running multiple agents in continuous loops burns tokens at a higher rate than single-prompt interactions. Steinberger recommended scheduling agents to work on hourly or daily cycles rather than running them continuously in the background, per The Economic Times. Osmani warned that additional sub-agents should only be deployed when the benefits justify the cost.
The practical implication: teams adopting loop engineering need to budget for AI compute the way they budget for human labor, with cost-per-task visibility and utilization tracking.
What This Shift Requires From Developers
Cherny’s prediction follows his earlier claim that software engineering itself is changing fundamentally. The progression is clear: if agents can write code, and agents can write the prompts that direct other agents to write code, the skill that matters is designing the loop architecture, not crafting individual prompts.
For developers building on platforms like OpenClaw, Claude, or Google Cloud’s agent tooling, the takeaway is concrete. The value is moving from “how do I phrase this prompt” to “how do I structure a system of agents that achieves this outcome autonomously.” That is a systems design problem, not a language problem.