Silicon Valley AI engineers are collectively shifting from prompt engineering to loop engineering.Article author and source: AI New Era
What's the hottest word in Silicon Valley lately?
Loop.
As soon as you open X, everyone online is talking about Loop Engineering.
Many big names in Silicon Valley are abandoning prompts and moving toward autonomous loops!
Huang Renxun has outlined the new focus for the next phase (and also a new way to burn tokens):
No one writes prompts anymore. The new job is to write and manage loops.
No one writes prompts anymore; the new job is writing Loops, working on Loops.
Recently, Anthropic engineers revealed:
Over 80% of Anthropic's engineers are already using self-improving loops, and this will reach 100% within 3 to 6 months.
Andrew Ng asserts: Prompts will die out in 3 to 6 months—Loop will replace prompts, and this is inevitable.
Previously, while explaining his AutoResearch project, Karpathy discussed the AI agent loop (generation → execution → evaluation → improvement) and called for partially removing humans from the loop.
Peter Steinberger, the creator of OpenClaw, bluntly says: "Monthly reminder—stop writing prompts by hand; designing loops is the way forward."
Boris Cherny, the creator of Claude Code, boldly declared: Loops are the future!
Two years ago, we were still writing code manually. Later, we began transitioning to having agents write code for us.
Now, we are moving toward a stage where agents prompt other agents to generate code.
Although the step from source code to agent is significant, the significance and impact of introducing a feedback mechanism are no less than those of the previous step.
Unexpectedly, after prompt engineering became popular for two years, top AI engineers have already moved on.
Why are Silicon Valley heavyweights all bullish on Loop?
The essence of traditional prompting is: humans are the loop itself.
You write the prompt → Agent outputs → You review → You write the next prompt → Repeat the cycle.
Each step relies on human attention, contextual memory, and decision bandwidth. The number of tokens and complexity of tasks one person can effectively manage in a day is limited.
The essence of Loop Engineering is: the system becomes the loop itself.
Therefore, loop engineering is more important than prompt engineering.
Humans only make one high-value design:
1. Define objectives and stop conditions
2. Set up a verification mechanism (most critical)
3. Establish persistent memory (Markdown / external state)
4. Configuration Discovery and Scheduling
Subsequently, the AI loop system can autonomously discover tasks → execute → verify → persist → and rediscover, operating 24/7, with human intervention only when necessary.
This is why Silicon Valley executives are all bullish on Loop.
They base this judgment on the belief that once the loop matures, the cost-effectiveness of manual prompts will plummet.
Loop Engineering's 11-page whitepaper is going viral across the internet.
So, what exactly is the looping process?
In recent days, a whitepaper on Loop Engineering has been going viral on X.
This 11-page PDF is essentially a popular summary or field guide that consolidates relevant public discussions and practical experience.
The core idea stems from public statements by Peter Steinberger, Boris Cherny, and Addy Osmani.
What is circular engineering?
Loop Engineering, named by Google Chrome engineer Addy Osmani in June 2026.
It is the fourth layer above prompt engineering, context engineering, and toolchain engineering: the first three layers assume you are seated at the keyboard, giving AI commands line by line; Loop Engineering removes you from that position entirely, liberating you from the role of doing the work.
From now on, you are no longer the engine driving the AI, but the architect designing the engine.
The system will automatically wake up at the scheduled time, initiate derivative sub-agents to perform tasks, and feed the output results back to itself as input for the next round.
The article breaks down a complete loop into five key actions:
Discovery: AI uses its fixed skill set to autonomously identify valuable tasks, such as reviewing the latest CI failure logs or unresolved issues.
Handoff: Launch a separate sandbox for each task to enable multiple agents to operate in parallel without interfering with each other.
Verification: This is the most critical step. An AI that writes code will only blindly praise itself. Therefore, a completely independent, default-skeptical "evaluator" agent must be introduced to identify errors.
Persistence: AI memory cannot remain solely within a context window that may be cleared at any time; its state and progress must be saved to disk so that it can resume where it left off the next day.
Scheduling: Automate the system to run periodically through scripts, closing the entire loop.
Among these, verification is the most difficult and the most easily overlooked or skipped.
When AI rates itself, it almost always praises itself because it carries a self-persuasion chain inside. The solution is to introduce an independent evaluation agent that assumes the code is flawed by default.
However, fully automated system operation does not mean you can rest easy. The author warns: when the cycle runs wild in the middle of the night, it may quietly accumulate four hidden costs.
Verify the debt: Unverified minor errors have been quietly merged into the ledger.
Understanding the rot: AI writes code too quickly, causing a severe disconnect between humans and the codebase.
Cognitive surrender: Humans stop reviewing and fully accept AI's results.
Token失控：AI在死循环中彻夜重试，烧光预算。
Two people working with the same loop may arrive at completely opposite outcomes. Bring in sound judgment, and it amplifies judgment; bring in laziness, and it amplifies laziness.
In summary, this report reveals a profound industry shift: circular engineering has made code generation nearly free, while human judgment has become the only scarce resource!
Additionally, a 14-step practical guide issued by Codez is also going viral across the network, with millions of shares so far.
The gist of the article is as follows: The prompt is outdated; the leverage point has shifted up one level—from “words written for AI” to “designing a system that automatically feeds AI.”
This transition can be broken down into 14 steps across three phases—
First, determine whether you truly need a loop (is the task repetitive? Can verification be automated? Can your budget afford it?), then master five core components (scheduling, isolated working directories, skill files, external connectors, independent evaluation sub-Agents), and finally build a minimal viable loop.
The most critical point is to separate the agent that writes the code from the agent that reviews it. The same model acting as both player and referee will always give itself a perfect score.
A loop without objective verification is merely "two optimists nodding at each other"—the better the loop performs, the more likely it is to cause engineers to stop truly understanding the code.
Loop Engineering Timeline
If we were to outline a timeline for circular engineering, it would be approximately as follows.
Initial foundational phase
2022: Yao Shunyu and others proposed the ReAct framework, laying the theoretical foundation.
2025: Geoffrey Huntley proposes "Ralph".
Early 2025–2026: Andrej Karpathy releases the AutoResearch project, establishing a classic autonomous experimentation loop—a major milestone.
Concept explosion and naming phase
In early June 2026, Peter Steinberger stated: "You should no longer manually prompt coding agents; instead, you should design loops that prompt them."
Boris Cherny said: I no longer directly prompt Claude; my job is to write the loop that runs Claude.
June 7, 2026: Addy Osmani published a blog post titled "Loop Engineering," formally naming it and introducing a four-layer stack framework: Prompt → Context → Harness → Loop Engineering.
Subsequently, throughout June, Loop Engineering went viral across the network.
Claude "Infinite Loop," automated agent takes over everything
In an internal podcast, Anthropic's engineers revealed a chilling detail:
When you click run and let Claude operate for 8 hours, you're essentially placing a $500 bet on computational power.
If you're still struggling with how to write prompts, you've already lost.
In Anthropic’s logic, engineers are evolving into “compute allocators.”
Your core work is no longer writing logic, but deciding where to allocate every penny of computing power.
As OpenAI researcher Noam Brown pointed out earlier this month, contemporary models can solve nearly any problem if you're willing to throw enough computational power at them.
Circuits engineering is part of the broader trend of "computation during testing."
Interestingly, the idea of making agents work in a loop has long had its early form.
At least last summer, Australian sheep farmer Geoffrey Huntley mentioned a similar approach on his blog, calling it the "Ralph Cycle."
A year ago, if you wanted to set up a loop, you had to write a bunch of bash scripts and forever maintain that code, which belonged only to you.
Now, these components are built directly into the product.
You would no longer argue over whether to use Codex or Claude Code, but instead design loops that work seamlessly across any tool you're using.
Addy Osmani, former head of engineering and developer relations at Google, noted that all the components required for circular engineering are already available in these AI tools.
He also asserted that circular engineering could lead engineers to "cognitive surrender":
When the cycle runs on its own, it’s easy to stop thinking and passively accept everything it feeds back.
This is also precisely what software engineer Armin Ronacher is concerned about.
This is exactly what the Orange Book says: Addy Osmani encourages everyone to design loops with discernment:
Designing routines can be a remedy when done with discernment, or a catalyst when done to avoid thinking—the same action, utterly opposite outcomes.
Endless cycling, engineering never dies
At the April AI Engineer Conference, Anthropic's engineers said they had Claude develop a retro-style mini-game app using two different approaches: one with minimal prompts only, and another using an agent loop.
The results are striking: the minimalist prompt version took 20 minutes and cost $9, while the iterative approach took 6 hours and cost $200.
But the latter produces an app with significantly higher quality.
The former game doesn't run properly, and the app is very basic; whereas the circular version is much richer, including many features the game designer intended.
The cycle does not pity those who give up thinking. It only turns your ignorance into debt in the code, at an even faster pace.
Those willing to maintain their clumsy understanding, continuously define the rules, and take responsibility for the final outcome will gain greater leverage in this paradigm shift than ever before.
Build the loop. Stay the engineer.
The prompt is dead; the loop is running.
And you are still the one who determines its meaning.