The next phase of AI-assisted software development may not involve people writing better prompts. Instead, developers could build autonomous systems in which artificial intelligence agents instruct, monitor and improve one another.

Anthropic co-founder and the creator of Claude Code, Boris Cherny, argues that manual prompting is beginning to give way to “loop engineering,” an approach where AI agents generate prompts and continue working towards a goal with limited human supervision.

“It’s an agent that prompts Claude. I don’t write the prompt anymore. Claude writes the prompt, and now I’m talking to that new Claude that is coordinating,” Cherny told Business Insider.

The shift marks an evolution from “vibe coding”, where developers describe what they want and allow AI tools to produce the code. Under a loop-based system, a user could provide a broad objective through a command such as `/goal`, after which agents would plan tasks, execute them, assess the results and continue iterating until the objective is met.

“You shouldn’t be prompting coding agents anymore. You should be designing loops that prompt your agents,” Peter Steinberger, creator of OpenClaw at OpenAI, told users.

Such systems could, for instance, use one agent to write code and another to independently review it. An agent could also periodically wake up, examine an active software repository, identify pending work and distribute tasks across separate threads.

Addy Osmani, a director at Google Cloud, said effective agent loops require five building blocks: automations, worktrees, skills, plugins and connectors, and sub-agents. Together, these components allow AI systems to access tools, divide work and operate continuously rather than wait for repeated instructions.

The developer’s role, as a result, could increasingly resemble that of a manager designing jobs and supervising teams.

“This is the time for the manager. You are designing a job,” ChatPRD founder Claire Vo said. “Just imagine you’re onboarding an employee — that employee could be an assistant, a customer service agent, or a software engineer.”

However, greater autonomy comes with higher computing costs. Networks of agents that repeatedly call models, review outputs and restart tasks can quickly consume large token budgets. Developers may need to decide whether tasks should run every few minutes, hourly or daily.

Osmani has also cautioned against deploying specialised sub-agents by default, arguing that the additional cost is worthwhile only when a task genuinely needs another opinion.