Creator of Claude Code Deletes IDE, Shifts to 'Loop Engineering': Prompts Not Dead, But Paradigm Has Changed

Boris Cherny, creator of Claude Code, deleted his IDE in November 2025 and announced in a June 2, 2026 interview that he no longer writes prompts, instead focusing on designing loops for AI to operate autonomously. Subsequently, OpenAI's Peter Steinberger and Google's Addy Osmani echoed the shift, with Osmani naming the new paradigm 'Loop Engineering'.

Core Shift: From Prompt Engineering to Loop Engineering

Over the past two years, developer interaction with AI coding agents has been turn-based: humans write prompts, read responses, then write the next prompt. Loop Engineering requires developers to build an automated system where AI autonomously discovers tasks, assigns execution, checks progress, and decides next steps. Osmani likens it to 'moving from operating a machine tool to designing an entire production line.'

Key Components of Loop Engineering

Osmani listed six components of the loop in a blog post:

Automations: Scheduled triggers that automatically discover and triage tasks.
Worktrees: Using git worktree to prevent conflicts between multiple agents.
Skills: Writing project conventions into SKILL.md to avoid agents relearning.
Connectors: Based on the MCP protocol, allowing agents to access issues, databases, Slack, etc.
Sub-agents: Separating code writing from review to avoid self-evaluation bias.
Memory: Using Markdown files to record progress and ensure continuity.

Practical Example: Boris's 'Hive' System

Developer @Av1dlive reconstructed a three-layer system called 'THE HIVE' based on Boris's public interviews:

Layer 1: Local loop (/loop), minimum 1 minute interval, session-level.
Layer 2: Cloud routines (Routines), minimum 1 hour interval, runs even after shutdown.
Layer 3: Cluster (/batch), runs hundreds of sub-agents in parallel without interference.

When combined, cloud routines discover tasks and write to files, local loops read and execute, and large tasks spin up the cluster. Results are distilled weekly into CLAUDE.md, enabling continuous agent evolution.

Prompts Are Not Dead

Loop Engineering does not eliminate prompts but embeds them within loops. Commands like /loop and /goal in Claude Code, and /goal and Automations in Codex, require developers to write prompts as loop instructions. Prompts have shifted from one-time inputs to reusable loop components, increasing their value.

Challenges and Controversies

Loop Engineering faces three major challenges:

Cost: Multi-agent parallelism consumes large token volumes; Steinberger joked that having 'unlimited tokens' is an OpenAI employee perk.
Loss of Control: Anthropic notes that loops are prone to laziness, self-praise, and drift, requiring separation of 'maker' and 'checker'.
Skepticism: Some developers argue that Loop Engineering is essentially a repackaging of old concepts like ReAct and AutoGPT, not a new paradigm.

Industry Impact

Codex and Claude Code are converging on loop components, with competition shifting to how easily users can write loops. Loop Engineering is seen as the center of a paradigm shift in AI programming, potentially marking the end of the editor era.