Huang Renxun: Prompt Is Becoming Obsolete, Loop Is the New Paradigm

NVIDIA CEO Jensen Huang recently proposed that AI programming is shifting from Prompt Engineering to Loop Engineering, where automated loop systems allow AI to iterate tasks autonomously rather than writing prompts one by one. This view has been echoed by AI leaders including "Father of Lobster" Peter, "Father of Claude Code" Boris Cherny, and Andrew Ng.

What is Loop

The core of Loop is to let AI autonomously run, verify, and iterate: users define goals, AI executes and checks results, retrying automatically if unqualified until passing or reaching a budget limit. Unlike Agent, which is an executor, Loop is a management mechanism that enables Agents to operate without continuous human intervention.

Product Implementation: Claude Code and OpenAI Codex

The Loop concept is already embodied in two major products:

• Claude Code: Provides /loop (timed loop), /goal (goal-driven), /schedule (cloud scheduled tasks). /goal adopts a "write-verify separation" design—large models write code, while a separate small model Haiku handles verification to avoid self-scoring bias.
• OpenAI Codex: Uses automated pipelines + goal-driven + multi-sub-Agent mode, running up to 8 Agents in parallel in cloud sandboxes, then aggregating results.

Both have different implementation paths but similar final forms: decompose complex tasks, execute with multiple Agents in parallel, and aggregate results. Community benchmarks show similar performance, indicating model capabilities are converging, with the gap lying in upper-level Loop orchestration.

Practical Guide and Pitfalls

X blogger Codez summarized 14-step practical tips:

1. 4-condition test: Is the task repetitive? Is there automated verification? Is the token budget sufficient? Does the Agent have advanced tools? Only build a Loop if all conditions are met.
2. Minimum Viable Loop: Includes trigger, skill (project context), state file (progress tracking), gate (auto-check). Order: manually run → write as Skill → wrap in Loop → schedule.
3. Write-verify separation: Use different models for writing and verifying to avoid lenient self-scoring.
4. Common pitfalls: Lack of hard stop conditions (set token/iteration/time limits), state not persisted (must write to file), having Loop handle judgment tasks (e.g., architecture rewrite, payment logic), not reading Diff leading to "understanding debt."
5. Core metric: Average cost per accepted change. If acceptance rate is below 50%, the Loop is losing money.

Paradigm Evolution: From Prompt to Loop

Although the Loop Engineering concept has emerged only in the past three weeks, the evolution path is clear:

• Prompt Engineering (2023-2024): Focus on prompt writing.
• Context Engineering: Optimize context provision.
• Harness Engineering: Design tools and frameworks.
• Loop Engineering: Build autonomous loop systems.

Boris Cherny's practice shows the ultimate form: since November 2024, he has uninstalled his IDE, with hundreds of Agents working in parallel, all code written by Claude Code, mostly on his phone. Agents prompt each other without human review.

Impact and Outlook

The Loop paradigm shifts the developer's role from "instruction giver" to "rule designer," promising to significantly boost AI programming efficiency. However, experts caution that Loop is suitable for tasks with clear right/wrong answers that machines can verify, while architectural decisions and payment logic still require human intervention.