Loop Engineering Rises: From Prompt Engineering to Loop Engineering, Silicon Valley Leaders Shift Focus

In June 2026, Loop Engineering, formally named by Google Chrome engineer Addy Osmani, quickly became the hottest concept in Silicon Valley. AI leaders including NVIDIA CEO Jensen Huang, Anthropic engineers, Andrew Ng, and Andrej Karpathy have all stated: Prompts are becoming obsolete; writing and managing loops is the new paradigm.

Core Concept: From "Writing Prompts" to "Designing Loops"

Traditional prompt engineering relies on humans writing instructions round by round, reviewing outputs, and then writing the next round, making humans the bottleneck of the loop. The essence of Loop Engineering is removing humans from the loop, allowing the system to autonomously complete the closed loop of "discover task → execute → verify → persist → discover again." The human role shifts from "instruction giver" to "rule designer," only needing to define goals, stopping conditions, verification mechanisms, and persistence states.

Key Principle: Separation of Execution and Verification

The core principle of Loop Engineering is "the agent that writes code and the agent that reviews code must be separate." When the same model acts as both player and referee, it almost always praises itself. Claude Code's product architecture embodies this principle: a large model writes code, while an independent smaller model, Haiku, handles verification. OpenAI Codex similarly uses multiple sub-agents executing in parallel with independent verification.

Practical Framework: Three Files to Build a Loop System

According to widely circulated tutorials, building a basic Loop requires just three files:

• Agent Definition: Define builder (responsible for writing and fixing code, with write permissions) and checker (only checks, with read-only permissions), achieving hard isolation at the tool level.
• Loop Orchestrator: Registered as a slash command (e.g., /loop), it schedules builder and checker to alternate, running up to 5 rounds, stopping if the same failure occurs twice consecutively.
• Stopping Rules: Written into the project root's CLAUDE.md, specifying hard stop conditions (e.g., token limit, iteration limit, time limit) to prevent runaway loops.

Product Deployment: Claude Code vs. OpenAI Codex

Current main product carriers for Loop Engineering include:

• Claude Code: Offers /loop (timed loop), /goal (goal-driven, until acceptance conditions are met), and /schedule (cloud-based scheduled tasks).
• OpenAI Codex: Uses automated pipelines + goal-driven + multiple sub-agents, supporting up to 8 agents working in parallel in cloud sandboxes.

Though their implementation paths differ, the final forms are highly similar—breaking complex tasks into pieces, distributing them to multiple agents for parallel execution, and aggregating results.

Industry Impact and Risk Warnings

Anthropic internal engineers reveal that over 80% of engineers are already using self-improving loops, with 100% adoption expected within 3-6 months. Andrew Ng asserts that prompts will die within 3-6 months.

However, fully automated operation brings four hidden costs:

• Verification Debt: Subtle unverified errors are quietly merged into the codebase.
• Understanding Decay: AI writes code too fast, causing human understanding of the codebase to fall significantly behind.
• Cognitive Surrender: Humans become too lazy to review, blindly accepting AI results.
• Token Runaway: AI retries all night in infinite loops, burning through budgets.

Community advice: Loops are suitable for tasks with "clear right/wrong and machine-verifiable" outcomes (e.g., auto-fixing lints, classifying CI failures), but not for areas requiring human judgment like architecture rewrites or payment logic. The only metric to measure loop effectiveness is the average cost per accepted change; if the acceptance rate is below 50%, the loop is losing money.