Breaking

Claude Code Lead Drops Manual Prompting for Self-Running AI Loops

July 4, 2026 at 02:32 EDT

AI Agents
Software Dev & Coding
Foundation Models

Boris Cherny, who leads Claude Code at Anthropic, has revealed that he has stopped manually prompting the AI and moved to designing "AI loops" that recursively call Claude to generate, test and improve code on their own. In several interviews he has said: "I don't prompt Claude anymore. I have loops running. They're the ones prompting Claude and figuring out what to do. My job is to write loops."

Anthropic · Head of Claude Code

"I don't prompt Claude anymore. My job is to write loops."

Boris Cherny says he stopped typing prompts into the AI and instead designs self-running loops that generate, run, and evaluate their own prompts — fueling a debate over a shift from prompt engineering to loop engineering.

100%

of his code handled by Claude Code — then he uninstalled his IDE

5–10+

Claude sessions run in parallel, automated via /loop

elements per loop: objective · metric · boundary

The self-running loop

Generate
loop writes the prompt

→

Run
executes in git worktrees

→

Evaluate
score against metric

↺ iterate automatically via cron · GitHub Actions · hooks

Paradigm shift

TRADITIONAL · PROMPTING

Unit: a single instruction
Human enters prompts each time
Manual, per-conversation runs
Challenge: effort of repetition

LOOP ENGINEERING

Unit: a persistent loop
Human designs objective · metric · boundary
Automated iteration via cron/hooks
Challenge: token use, need for oversight

The upside

"Improvements run while you sleep"
System-design skills matter more
Parallel persona agents raise bug detection

The cautions

Token use surges; plan limits hit sooner
A weak metric breeds a "bad loop" of low-quality output
Human oversight still essential — "just an evolved cron job?"

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.