Following the debut of ChatGPT in 2022, the way people interacted with artificial intelligence (AI) revolved around one simple idea: Users write a prompt, share enough context, and the AI model generates a response. In the process, prompt engineering quickly emerged as a coveted skill. The better the prompt, the better the output.
For a time, prompting was the interface. Then came AI agents, capable of carrying out tasks autonomously with a little hand-holding from human users. Users can assign tasks, review the agent’s work, and step in when needed to provide guidance or approval.
Now, that workflow is also beginning to evolve. Developers are setting up recurring systems with a defined purpose that the AI agents iterate until complete. This is known as loop engineering. It eliminates the need for a user to type in a prompt using their fingers in order for the AI agent to work on their behalf.
Several AI experts have urged more developers to embrace loop engineering. Boris Cherny, the head of Anthropic’s Claude Code, said that he does not write AI prompts anymore and only relies on AI agent loops
“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 kind of coordinating,” he was quoted as saying by CNBC. The Anthropic executive further said that loops and a similar feature were examples of the kind of work he would be proudest of in a decade.
Peter Steinberger, an OpenAI engineer and creator of the viral OpenClaw project, recently addressed users who are still writing out prompts for AI agents. “Here’s your monthly reminder that you shouldn’t be prompting coding agents anymore. You should be designing loops that prompt your agents,” Steinberger wrote in a post on X.
Stating that the days of directly prompting AI coding tools were “kind of over”, Addy Osmani, director of Google Cloud, wrote, “Loop engineering is replacing yourself as the person who prompts the agent. You design the system that does it instead […] I believe this may be the future of how we work with coding agents.”
What are AI agent loops?
Simply put, a loop is used to describe a recurring system that guides AI agents, so the user does not have to constantly write prompts themselves. As Osmani puts it: “You let that system poke the agents instead of you.”
Each loop comprises five key components, according to the Google executive:
-Automations: They are the foundation of a loop and ensure that a loop can be repeated rather than a one-off event. It enables the AI agent to go off on a schedule and do discovery and triage by themselves.
-Worktrees: This is a key component as they enable two AI agents to work in parallel with each other, avoiding any overlap.
-Skills: They can, for instance, be instructions for the AI agent to write down the project knowledge the agent would otherwise just guess.
-Plugins and connectors: They are used to give the AI agent access to the tools you already use.
-Sub-agents: They let users set up AI agents in such a way that one of them has the idea while another one checks the work.
Osmani also mentioned memory as a sixth component, and advised developers to store information such as what is done, what is next, etc, in a markdown file or Linear board. He said that “the model forgets everything between runs so the memory has to be on disk and not in the context.”
How are loops used by developers?
There are several examples of AI agent loops set up by developers to eliminate the need for prompting at each step. For instance, /goal is an example of a loop that instructs an AI coding agent like Claude Code or OpenAI’s Codex to keep working until a task is completed.
OpenClaw’s Steinberger said that he has implemented a Codex loop which instructs the AI tool to maintain repositories, wake up every 5 minutes, and direct work to threads. “That makes it easy to parallelise+steer work as needed,” he said. Experts have also urged developers to split up their loop in such a manner that one agent writes the code and the other checks the final product. It is not advisable to have the same model that wrote the code review its own work because it could be “way too nice” in its evaluation, as per Osmani.
Not all AI agent loops have to be related to coding. Claire Vo, host of the ‘How I AI’ podcast, suggested that managers can also design loops to automate, for instance, the onboarding process of employees. “And so just imagine that you’re onboarding an employee […] You’re going to say, ‘Every Friday, EA, I would like you to review my calendar, see who cancelled on me, where I could have used my time effectively, if there are any followups, and send me a Slack to get this done. And I want you to do that every Friday’,” Vo said.
What are the challenges?
If your loops involve running multiple agents with sub-agents powered by the latest frontier AI model, you would probably have to increase your personal token budget by a lot.
Loops are known to burn through lots of tokens and rack up costs for users because they often involve long-running tasks that span many steps and do not always have an end-point.
When asked about how he would modify his loop to be more budget-conscious, Steinberger wrote, “Waking up and doing some API calls is fairly cheap, or opt for once per hour/day for lower token use.” However, the OpenClaw creator also joked that he is someone who has unlimited tokens given that he now works for OpenAI.
Advising developers to spend only where necessary, Osmani said, “Subagents do burn more tokens since each one does its own model and tool work, so spend them where a second opinion is worth paying for.”
Claude Code users could also opt for Scheduled Tasks instead of running the AI coding agent in Loop mode. Scheduled Tasks are considered to be better for cost-senstive workflows because they only run at specific times as opposed to Loop mode which is a continuous, iterative process.
Hence, using an AI agent to run a task for five minutes a day consumes far less tokens than having the agent run on a loop all day looking for something to do.
On the security front, Osmani also warned against relying entirely on automated loops and said that human prompting of agents can also be effective. “Build the loop. But build it like someone who intends to stay the engineer, not just the person who presses go,” he wrote.
Loops are known to burn through lots of tokens and rack up costs. (Image used for representational purposes/Unsplash) 
Share
