Nobody knows but him and Boris
On June 7, Peter Steinberger posted a reminder that cleared 2.2 million views and dominated the AI-coding conversation for a week.
"Here's your monthly reminder that you shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents." - @steipete, June 7, 2026
Varadh Jain asked what that looks like in practice, and Matthew Berman's answer became the most quoted reply: "nobody knows but him and boris."
The reply stuck because it was nearly true: one camp declared prompt engineering dead, another called loops a cron rebrand, and few on either side could define the thing they were arguing about. Boris Cherny had defined it on stage five days earlier.
What Boris means by a loop
Cherny started Claude Code as a side project in September 2024, and it now sits behind close to 4 percent of public commits on GitHub. At the WorkOS Acquired Unplugged event he described his current workflow.
"I don't prompt Claude anymore. I have loops that are running. They're the ones that are prompting Claude and figuring out what to do. My job is to write loops." - Boris Cherny, June 2, 2026
A loop is a small program that prompts the agent, reads what it produced, checks whether the task is done, and prompts again if it isn't. You write the loop once and the model runs inside it as a subroutine.
His own numbers back it: in the 30 days before December 27, 2025, every one of his contributions to Claude Code was written by Claude Code, 259 PRs in total. He deleted his IDE in November and has not opened one since.
The five-year lineage
The 2022 ReAct paper formalized the first version: the model reasons, calls a tool, reads the result, and repeats until done. AutoGPT in 2023 let the model prompt itself toward a goal and became famous for spinning forever.
Geoffrey Huntley's ralph loop, published in July 2025, pipes the same prompt file into the agent over and over and resets the context to fixed anchor files each pass. He built a working programming language with it for about 297 dollars.
In spring 2026 both Codex and Claude Code shipped that pattern as /goal, which reruns the task until a small validator model confirms it is done.
The new layer, and the one Steinberger means, is loops that supervise other loops. They run on a schedule instead of a human kickoff and keep state in git so they survive a restart.
Steve Yegge's Gas Town, open-sourced in January, is the built version: 20 to 30 Claude Code instances coordinated by a Mayor agent, with patrol agents running continuous loops.
Cron with a decision in it
The best skeptic reply was four words: "Cronjobs have funny re-branding rn." It is half right: the scheduling layer is cron, and Claude Code's /loop uses cron under the hood.
The difference is the body of the job: cron runs a fixed script, while a loop runs a model that reads the current state, picks the next action, and checks whether it worked.
Let one loop dispatch and supervise others over shared durable state and you have a structure cron cannot express.
The bill moved to the loop
The engineers shipping agents were blunter about what a loop is.
"Every ai agent i shipped this year is a for-loop, an llm call, and a try/catch around the json parsing. The only thing agentic about it is the anthropic bill at the end of the month." - @rohit_jsfreaky, June 2026
The bill is real: Uber capped its engineers at 1,500 dollars per person per tool per month for Claude Code and Cursor after burning its annual AI budget in four months. The model writes the code for close to nothing, so the cost moved to running the loop.
The failure everyone fears in production is the loop that never halts, which is why serious setups converge on a maximum iteration count, no-progress detection, and a token or dollar ceiling.
Gartner puts agentic AI at the peak of inflated expectations, with about 17 percent of organizations actually deploying agents.
The answer Berman wanted
Cherny posted his own starter loop days after the tweet, and it answers Jain's question in one line:
/loop babysit all my PRs. Auto-fix build issues, and when
comments come in, use a worktree agent to fix them.You write the intent and the stopping behavior, and the loop prompts the agent each tick. Cherny's five tips for running Opus unattended end with the one that decides whether the loop is useful: make sure the agent can verify its own work end to end.
A loop with no check accepts output that looks finished whether or not it runs, which is why Dan Kornas's roborev reviews every commit in the background and feeds the findings back to the agent while the context is still fresh.
Steinberger's companion advice pairs with the tweet: anything you do more than once, turn into a named skill the loop can call, so each run starts from tested pieces instead of re-deriving them.