Boris Cherny Says He Stopped Writing Prompts and Started Writing Loops

Boris Cherny runs Claude Code at Anthropic. In a 40-minute fireside chat at Meta’s Scale conference that surfaced on June 22, 2026, he told the audience he had stopped typing prompts and now designs autonomous systems that prompt Claude for him, according to the 40-minute Meta Scale fireside chat.

The video spread across developer timelines within hours. It also crystallised a shift several senior engineers had been circling in the same week. Peter Steinberger, the creator of OpenClaw, posted on June 7, 2026 that “you shouldn’t be prompting coding agents anymore. You should be designing loops that prompt your agents.” Addy Osmani published a technical breakdown of the same idea days later. Three senior engineers at three different organisations landed on almost identical language in seven days.

Why the Head of Claude Code Stopped Writing Prompts

Boris Cherny created Claude Code as a side project in September 2024 and now leads the tool inside Anthropic. In the Meta Scale fireside chat, he framed the change as personal and ongoing, not speculative.

Cherny said “100% of my code has been written by Claude Code” since Opus 4.5, and added that most of it now happens straight from his phone. He no longer types instructions into a chat box. He designs systems that type instructions into the model on a cadence, then check whether the model succeeded. The phrase that caught on is “my job is to write loops.”

That redefinition puts the engineer outside the conversation. The model is now a subroutine called by a loop, and the engineer authors the loop. Claude Code reportedly sits behind close to four percent of all public commits on GitHub.

What a Loop Actually Is

A loop is a small autonomous system that prompts an AI agent, reads what it produced, decides whether it is done, and if not, prompts it again. The decision is the agent’s, not a hardcoded branch.

One blunt restatement of the same idea circulated on X in the same week: “cronjobs have funny re-branding rn.” That line is half right. Cron never had a model in the middle, deciding what to do next based on the current state of the work. Stack the model, the scheduler, the verifier, and a place to write down what is done, and you get something plain cron cannot express. A loop is a schedule plus a decision-maker in the body of the run.

The comparison that sharpens the idea is between a prompt and a loop. A prompt is a one-shot instruction that produces one output and waits for the human. A loop is a continuous operation that produces outputs until a verifiable condition is met, with no human turn required between cycles. Claude Code’s own primitives make the difference concrete: /loop re-runs on a cadence, and /goal keeps going until a small validator model confirms the task is done.

A useful way to read this is that the unit of work changed. Two years ago, the unit of work was a turn, one prompt in, one response out. In 2026, the unit of work is a loop, an autonomous system that runs on a schedule and owns one slice of work until it is done. The shift is the same kind of move that took us from running one program at a time to running dozens of services on a scheduler. Cherny, Steinberger, and Osmani all reached the same framing without coordinating, and how loops are built in Claude Code and Codex maps onto the same primitives in both tools.

How Codex and Claude Code ship the same loop primitives:

The Six Pieces Every Loop Needs

Most accounts of loops converge on five capabilities plus one piece of state. Addy Osmani’s technical breakdown is the most detailed. The five are automations, worktrees, skills, plugins and connectors, and sub-agents; the sixth is memory, a file or board that lives outside the conversation and survives across runs.

The list is the same in Claude Code and OpenAI’s Codex, with only the names changing. Automations are the heartbeat, triggering discovery and triage on a schedule. Worktrees isolate parallel agents so two of them writing the same file do not collide. Skills are reusable SKILL.md files, connectors plug the agent into GitHub, Slack, and Linear via MCP, and sub-agents split the worker from the verifier.

• Automations that trigger on a schedule, the heartbeat of the loop
• Worktrees so parallel agents do not collide on the same files
• Skills, reusable SKILL.md files that codify project knowledge
• Connectors built on MCP that link the loop to GitHub, Slack, Linear, and the rest of the stack
• Sub-agents, one to explore or implement, a separate one to verify
• Memory, a state file or board that survives across runs

Three Engineers Converged on the Same Idea in One Week

Peter Steinberger’s June 7, 2026 tweet cleared 2.2 million views and triggered a reply chain arguing over what a loop even is. Matthew Berman replied that nobody knows what loops look like in practice “but him and boris.”

That reply captured the state of the discourse. The phrase hit two million views, and the people boosting it were still arguing about what it meant. Addy Osmani published a long technical breakdown on his blog in the same window, mapping the five primitives onto both Claude Code and Codex with examples.

Cherny has been making these points across multiple recent appearances, including the Acquired podcast earlier in June and the Meta Scale talk that landed on June 22. When three senior engineers from three different organisations describe the same change in the same week, using the same word, the word is no longer a buzzword. It is the label the industry has settled on for the new shape of work.

The convergence is not a coincidence. Cherny, Steinberger, and Osmani are all watching the same product cycle. Claude Code and Codex both shipped /goal in spring 2026, the primitive that runs the loop until a small validator confirms the task is done. Geoffrey Huntley’s ralph loop, a bash one-liner from July 2025 that pipes the same prompt file into the agent repeatedly, became the pattern product teams built on top of. The primitives had to be in the products before anyone could name the pattern, so the naming took until both tools shipped them.

One way to see how durable the shift is: Steve Yegge’s Gas Town coordinates twenty to thirty Claude Code instances with a Mayor agent and continuous patrol loops, with state stored in git so a crash does not lose work. The scale is unusual. The shape is not. Anyone running Claude Code today can copy the primitives in an afternoon.

The size of the shift, in numbers:

• 2.2 million views on Peter Steinberger’s June 7, 2026 tweet about loops
• 20 to 30 Claude Code instances coordinated in Steve Yegge’s Gas Town, with state stored in git
• $1,500 per person per tool per month cap Uber set for Claude Code and Cursor after four months of burn
• 4 percent of all public commits on GitHub Claude Code reportedly sits behind, per a June 2026 analysis

The Hard Part Has Moved from Writing to Verifying

Boris Cherny did not just say he stopped prompting. He also said what takes up his time now. The verification layer is where the loop either holds together or quietly breaks. A loop is only as trustworthy as its ability to check its own work.

I don’t prompt Claude anymore. I have loops running that prompt Claude and figuring out what to do. My job is to write loops.

Boris Cherny, head of Claude Code at Anthropic, in a 40-minute fireside chat at Meta’s Scale conference in June 2026.

The structure that works is a maker and a checker split. One agent, or a small team, explores and implements. A different one, sometimes a stronger model, always with different instructions, verifies against specs and tests. Claude Code’s /goal applies the same split to the stop condition itself. After every turn, a separate small model checks whether the loop is done, so the worker is not grading its own homework. Without it, the loop cannot stop on its own.

The cost of getting verification wrong is concrete. A loop with no maximum iteration count, no-progress detection, or dollar budget will burn through the cap before breakfast. Uber reportedly capped its engineers at $1,500 per person per tool per month for Claude Code and Cursor after burning through its annual AI budget in four months.

Where Loops Are Already Showing Up Outside Coding

Loop thinking is moving beyond engineering. A content team can monitor social channels every 30 minutes and surface trending topics; a startup ops lead can flag support ticket spikes the moment they happen.

The common shape is the same. A scheduled trigger scans a source, a connector pulls the data into a file the agent can read, and a sub-agent classifies the findings while a second one verifies. The loop writes its findings to a state file or a Linear board. Anything it cannot handle lands in a triage inbox for a human. Addy Osmani writes that loops are early, the economics are real, and you have to be careful about token costs.

Frequently Asked Questions

What is a loop in AI coding?

A loop is a small autonomous system that runs on a schedule, prompts an AI coding agent, reads the result, decides whether the task is done, and either stops or runs again. The decision is made by a model in the body of the run, not by a hardcoded branch. The loop persists state to a file or board outside the conversation so the next run does not start from zero.

How is a loop different from a prompt?

A prompt is a one-shot instruction that produces one output and waits for a human to type the next one. A loop is a continuous operation that produces outputs until a verifiable condition is met, with no human turn required between cycles. Claude Code’s /loop runs on a cadence; /goal runs until a separate small model confirms the work is done. The difference matters because a loop compounds across runs and a prompt does not. A prompt’s leverage is the cleverness of the words; a loop’s leverage is the system that runs the words on a schedule.

What are the core building blocks of a loop?

Five capabilities and one piece of state. Automations trigger on a schedule; worktrees isolate parallel agents. Skills are SKILL.md files that codify project knowledge; connectors on MCP link the loop to GitHub, Slack, and Linear. Sub-agents split the worker from the verifier, and memory holds the state across runs.

What is the biggest risk in running a loop?

Verification. The agent that wrote the code is too generous grading its own homework, so well-designed loops split a maker and a checker into separate agents, sometimes with a stronger model for the verifier. Without a maximum iteration count, no-progress detection, and a token or dollar budget, a loop will burn through a monthly AI cap before breakfast; Uber reportedly hit that wall after four months and capped engineers at $1,500 per person per tool per month.