Jensen Huang: AI Paradigm Shifting from Prompt to Loop, Humans Exiting the Execution Chain
0xBroomberg
Nvidia CEO Jensen Huang declared AI engineering has entered the Loop era — humans no longer give AI step-by-step instructions but design mechanisms that let AI cycle autonomously, self-correct, and keep running, marking AI's formal shift from passive tool to self-driving system.
What is a Loop, and how does it fundamentally differ from a Prompt?
In the Prompt era, a human wrote one instruction, AI returned one answer — every step required a human push.
A Loop flips this: the human defines only the goal. AI executes, checks its own output, and if it falls short, carries the error information into a new iteration — repeating until the task passes or the budget runs out.
This means → the human role shifts from "controlling AI step by step" to "designing rules that let AI run on its own." A human no longer needs to be present in the execution chain.
Is this just Huang's view, or an industry consensus?
It goes well beyond Huang. Boris Cherny, a core developer of Claude Code, has uninstalled his traditional IDE entirely. He now runs hundreds of small Agents in parallel — each scanning GitHub Issues, reading Slack feedback, or monitoring CI failures.
In plain terms = his workflow today is: AI does the work; only problems AI cannot handle land in his inbox. He says all his code has been generated by AI since Claude Opus 4.5 launched, and most of his work happens on his phone.
Andrew Ng, ReAct-framework creator Shunyu Yao, and other leading figures have also publicly adopted Loop-based workflows.
Which products are leading, and how do they work?
Loop engineering has converged into a Claude Code vs. OpenAI Codex duopoly.
Claude Code ships a three-part toolkit: `/loop` for timed cycling, `/goal` for objective-driven runs, `/schedule` for cloud-based scheduled tasks. The key design in `/goal` is separating code-writing from acceptance testing — a large model writes the code; an independent smaller model, Haiku, validates it, preventing the model from grading its own work.
OpenAI Codex takes an "automated pipeline" approach. In developer tests, up to 8 Agents run simultaneously in separate cloud sandboxes, with results merged at the end.
Two paths converging — what does that tell us?
The two differ in architecture, but their end state is remarkably similar: decompose a complex task, assign pieces to multiple Agents running in parallel, then unify the results.
This means → the capability gap between models is narrowing fast. The real competitive moat has shifted to Loop orchestration and systems engineering on top of the model layer.
In plain terms = the race is no longer about who has the smartest model — it is about who can organize a swarm of AIs most effectively.
From Prompt to Loop — how many paradigm shifts has AI engineering gone through?
2023–2024: Prompt Engineering — the core question was "how to ask AI."
2024–2025: Context Engineering — the question upgraded to "what materials to feed AI." 2025–2026: Harness Engineering — building runtime environments where AI can do real work.
2026–present: Loop Engineering — designing self-sustaining cycles that keep AI systems running autonomously. This reflects a consistent pattern: with each shift, the human's control grain moves higher — from "writing a sentence" to "designing a self-driving loop."
What does this mean for investors?
The Loop concept has been circulating in industry for less than three weeks, but the underlying technical lineage — the ReAct framework and its successors — has been building for years. This is not built on thin air.
This means → as the capability gap between foundation models keeps shrinking, whoever builds the strongest systems-engineering edge at the Loop orchestration layer will hold the advantage in the next phase of AI competition.
For investors, the focus should shift from "whose model scores highest on benchmarks" to "whose AI systems engineering is strongest."
Content is for reference only, not financial advice.