Loop Engineering: A Paradigm Shift in AI Workflows
The traditional model of AI application is typically a serial human-computer interaction: a user asks a question, the AI generates an answer, the user reviews it and suggests changes, and the AI revises. The entire process’s efficiency is capped by the speed of human feedback. The AI’s computational power lies largely idle while waiting for human thought, judgment, and instruction.
In June 2026, a concept called “Loop Engineering” gained widespread attention in the AI development community, with its core objective being to break this bottleneck. The idea was named by Google’s Addy Osmani and is actively promoted by engineers like Boris Cherny, head of Claude Code at Anthropic. Its central principle is to remove humans from the repetitive feedback cycle and replace it with an automated “generate-review” loop. This system is not accomplished by a single AI but must include at least two functionally separate AIs: one as the executor (generating content) and another as the reviewer (evaluating quality).
Anthropic’s internal research found that a single AI tends to be overconfident when evaluating its own work, making it difficult to spot deeper flaws in its output. Introducing an independent reviewer AI is equivalent to building an objective “second opinion” into the system, which is crucial for ensuring output quality and achieving true automation.
Architecture Breakdown: Building an Autonomous Review System
An effective Loop Engineering system typically includes several key components. Its architecture can be understood through a tutorial shared by Datawhale community member “Xiao Ke,” which uses code checking as an example to propose a clear “builder-checker” model.
- Dual Agents: The system includes at least one “Analyst” or “Builder” AI responsible for core tasks like data processing, code writing, or report generation. It also requires a “Reviewer” or “Checker” AI dedicated to evaluating the former’s output against predefined criteria.
- Tool Isolation: To ensure the independence and security of the review, the Reviewer AI’s permissions should be strictly limited, for example, granting it read-only access and prohibiting it from directly modifying the Analyst’s work. All modifications must go through the feedback loop and be completed by the Analyst itself.
- Loop Orchestrator: A central controller manages the entire workflow. It receives the Analyst’s output and submits it to the Reviewer. The Reviewer returns its assessment. If the output fails, the Orchestrator sends the assessment report with the identified issues back to the Analyst for the next iteration.
- Stopping Conditions: The loop cannot run indefinitely. Clear stopping conditions must be preset, such as reaching a maximum number of iterations (e.g., 5) or the Reviewer marking all checklist items as “Pass.”
- Escalation Protocol: When the loop reaches its maximum iterations without passing, or encounters an unresolvable error, the system needs an escalation mechanism to “report” the problem to a human user for a final decision.
Case Study: Automated Analysis of ‘Crowded Trades’ in AI Hardware Stocks
To validate this model’s effectiveness, an investment research project based on Loop Engineering was implemented. The task was to analyze the “crowded trade” phenomenon in the AI hardware sector and compare it with historical examples like liquor and new energy stocks, using data from the Tushare financial data API spanning from 2018 to the present.
In this experiment, human intervention was strictly limited to two instances:
First Human Intervention: Rule-Setting. At the start of the task, the user set clear analysis objectives and defined key “honesty principles”: all data comparisons must be on an apples-to-apples basis, data sources must be traceable, and any data fabrication would lead to immediate termination of the task.
Subsequently, the system, composed of an “Analyst” AI and a “Reviewer” AI, began to operate autonomously. In the first round, the Analyst AI successfully extracted data from Tushare, selected 14 core AI hardware stocks from areas like optical modules, PCBs, and servers, and generated a report. The report stated that the sector had surged by 508% in the past two years, far outpacing the CSI 300 Index’s 40% gain over the same period.
However, when comparing with historical hot sectors, the Analyst AI used an unfair basis to highlight its conclusion: it pointed out that the AI hardware stock pool’s trading volume accounted for 6.53% of the total market, much higher than the 2.13% at the peak of the liquor stock craze. The Reviewer AI, following the preset “fair comparison” principle, immediately identified the issue—the former figure was based on the trading volume of 14 stocks, while the latter was based on only 4. The different bases made the comparison of absolute values unconvincing.
The Reviewer AI fed this issue and specific details back to the Analyst AI. In the second iteration, the Analyst AI corrected its methodology, using the more equitable Shenwan industry classification for comparison. It pointed out that the current combined trading volume of the Electronics and Communication sectors accounted for 33.8% of the total market, significantly higher than the historical peaks of Food & Beverage (6.9%) and Electrical Equipment (16.0%). This conclusion was equally powerful and logically sound. The Reviewer AI also verified the accuracy of details like unit conversions.
Second Human Intervention: Final Acceptance. After the Reviewer AI gave a green light to all aspects of the second-round report, the system stopped the loop and delivered the result. The human user only needed to confirm this final report, which had already undergone multiple rounds of internal review.
The New Role of Humans: From Executor to System Architect
Loop Engineering demonstrates that AI automation does not eliminate the human role but rather elevates it from tedious, micro-level task supervision to a more macro, decisive position.
In this new paradigm, the core responsibility of humans is no longer to review line-by-line and provide feedback, but to become the system’s “architect” and “legislator.” The success or failure of the work heavily depends on whether humans can clearly define goals, set bottom lines, and establish fair evaluation criteria before launching the task. The reason the Reviewer AI could spot the Analyst’s “clever shortcuts” was that humans had pre-installed principles of “honesty” and “fairness” into its judgment criteria.
AI can efficiently generate solutions, but it lacks value judgment and a sense of ultimate purpose. After outsourcing workflows to AI, humans must retain the ability to define and safeguard these core principles. Otherwise, when faced with a perfectly crafted report from an AI but unable to identify its deeper fallacies, the true bottleneck will no longer be the AI, but the degradation of human cognitive abilities.
