Following the debut of ChatGPT in 2022, the way in which folks interacted with synthetic intelligence (AI) revolved round one easy concept: Customers write a immediate, share sufficient context, and the AI mannequin generates a response. Within the course of, immediate engineering shortly emerged as a coveted ability. The higher the immediate, the higher the output.
For a time, prompting was the interface. Then got here AI brokers, able to finishing up duties autonomously with slightly hand-holding from human customers. Customers can assign duties, assessment the agent’s work, and step in when wanted to supply steerage or approval.
Now, that workflow can be starting to evolve. Builders are establishing recurring programs with an outlined objective that the AI brokers iterate till full. This is called loop engineering. It eliminates the necessity for a person to sort in a immediate utilizing their fingers to ensure that the AI agent to work on their behalf.
A number of AI consultants have urged extra builders to embrace loop engineering. Boris Cherny, the top of Anthropic’s Claude Code, stated that he doesn’t write AI prompts anymore and solely depends on AI agent loops
“It’s an agent that prompts Claude. I don’t write the immediate anymore. Claude writes the immediate, and now I’m speaking to that new Claude that’s type of coordinating,” he was quoted as saying by CNBC. The Anthropic government additional stated that loops and an analogous function have been examples of the type of work he could be proudest of in a decade.
Peter Steinberger, an OpenAI engineer and creator of the viral OpenClaw mission, just lately addressed customers who’re nonetheless writing out prompts for AI brokers. “Right here’s your month-to-month reminder that you simply shouldn’t be prompting coding brokers anymore. You ought to be designing loops that immediate your brokers,” Steinberger wrote in a put up on X.
Right here’s your month-to-month reminder that you simply shouldn’t be prompting coding brokers anymore.
You ought to be designing loops that immediate your brokers.
— Peter Steinberger 🦞 (@steipete) June 7, 2026
Stating that the times of immediately prompting AI coding instruments have been “type of over”, Addy Osmani, director of Google Cloud, wrote, “Loop engineering is changing your self as the one who prompts the agent. You design the system that does it as an alternative […] I imagine this can be the way forward for how we work with coding brokers.”
What are AI agent loops?
Merely put, a loop is used to explain a recurring system that guides AI brokers, so the person doesn’t should continuously write prompts themselves. As Osmani places it: “You let that system poke the brokers as an alternative of you.”
Story continues under this advert
Every loop includes 5 key parts, in response to the Google government:
-Automations: They’re the inspiration of a loop and make sure that a loop will be repeated quite than a one-off occasion. It allows the AI agent to go off on a schedule and do discovery and triage by themselves.
-Worktrees: It is a key element as they permit two AI brokers to work in parallel with one another, avoiding any overlap.
-Abilities: They’ll, as an illustration, be directions for the AI agent to write down down the mission data the agent would in any other case simply guess.
Story continues under this advert
-Plugins and connectors: They’re used to present the AI agent entry to the instruments you already use.
-Sub-agents: They let customers arrange AI brokers in such a method that considered one of them has the concept whereas one other one checks the work.
Osmani additionally talked about reminiscence as a sixth element, and suggested builders to retailer data resembling what is completed, what’s subsequent, and so forth, in a markdown file or Linear board. He stated that “the mannequin forgets all the pieces between runs so the reminiscence must be on disk and never within the context.”
How are loops utilized by builders?
There are a number of examples of AI agent loops arrange by builders to remove the necessity for prompting at every step. As an illustration, /aim is an instance of a loop that instructs an AI coding agent like Claude Code or OpenAI’s Codex to maintain working till a process is accomplished.
Story continues under this advert
OpenClaw’s Steinberger stated that he has carried out a Codex loop which instructs the AI software to keep up repositories, get up each 5 minutes, and direct work to threads. “That makes it simple to parallelise+steer work as wanted,” he stated. Consultants have additionally urged builders to separate up their loop in such a fashion that one agent writes the code and the opposite checks the ultimate product. It’s not advisable to have the identical mannequin that wrote the code assessment its personal work as a result of it might be “method too good” in its analysis, as per Osmani.
Not all AI agent loops should be associated to coding. Claire Vo, host of the ‘How I AI’ podcast, prompt that managers may design loops to automate, as an illustration, the onboarding strategy of staff. “And so simply think about that you simply’re onboarding an worker […] You’re going to say, ‘Each Friday, EA, I would love you to assessment my calendar, see who cancelled on me, the place I might have used my time successfully, if there are any followups, and ship me a Slack to get this executed. And I need you to do this each Friday’,” Vo stated.
What are the challenges?
In case your loops contain working a number of brokers with sub-agents powered by the newest frontier AI mannequin, you’d in all probability have to extend your private token funds by lots.
Loops are identified to burn by way of plenty of tokens and rack up prices for customers as a result of they usually contain long-running duties that span many steps and don’t all the time have an end-point.
Story continues under this advert
When requested about how he would modify his loop to be extra budget-conscious, Steinberger wrote, “Waking up and doing a little API calls is pretty low cost, or go for as soon as per hour/day for decrease token use.” Nevertheless, the OpenClaw creator additionally joked that he’s somebody who has limitless tokens on condition that he now works for OpenAI.
Advising builders to spend solely the place mandatory, Osmani stated, “Subagents do burn extra tokens since each does its personal mannequin and power work, so spend them the place a second opinion is price paying for.”
Claude Code customers might additionally go for Scheduled Duties as an alternative of working the AI coding agent in Loop mode. Scheduled Duties are thought of to be higher for cost-senstive workflows as a result of they solely run at particular instances versus Loop mode which is a steady, iterative course of.
Therefore, utilizing an AI agent to run a process for 5 minutes a day consumes far much less tokens than having the agent run on a loop all day searching for one thing to do.
Story continues under this advert
On the safety entrance, Osmani additionally warned towards relying fully on automated loops and stated that human prompting of brokers may also be efficient. “Construct the loop. However construct it like somebody who intends to remain the engineer, not simply the one who presses go,” he wrote.