Woofun AI reports that Wen Ming, founder of API7.ai, spent billions of Tokens in a single month to utilize AI agents for rewriting a production-grade gateway, concluding that the industry bottleneck has shifted from AI capability to human cognitive limits. This initiative, which began with a specific debugging incident during the Spring Festival of 2026 involving Apache APISIX, evolved into a comprehensive reconstruction of the system's architecture, codebase, testing protocols, and documentation. The core finding is that AI can now execute coding tasks in half the time previously required for a three-month project, effectively matching or surpassing the output velocity of senior engineers, yet the strategic direction remains entirely dependent on human judgment.

The catalyst for this transformation occurred when the team encountered a persistent, non-reproducible bug within Apache APISIX that resisted traditional debugging methods. After exhausting manual code reviews and attempting various reproduction scenarios without success, the team resorted to describing the symptoms to an AI agent. Within ten minutes, the agent performed static code analysis and accurately identified the error location, a feat that had eluded human engineers for days. This event demonstrated that AI could handle complex diagnostic tasks with a speed and precision that manual intervention could not match. Consequently, the team questioned whether AI could manage the entire lifecycle of a production-grade project, leading to the decision to recreate the system from scratch. While a full recreation in a short timeframe was initially deemed impossible, the team discovered that by clearly defining the architecture, technical stack, and underlying conceptual relationships, AI could execute the actual coding phase in half the time of the traditional three-month cycle.

Wen Ming emphasizes that the current limitation is not the AI's ability to generate code, but the human capacity to keep pace with the volume of decisions required. In the past month alone, the personal expenditure on Tokens for software development reached the scale of billions, reinforcing the observation that AI capabilities have far exceeded practical constraints. The value of a senior engineer has fundamentally shifted; it is no longer defined by the ability to write syntax, but by the accumulation of mistakes, technical trade-offs, and architectural insights gained over years of experience. These critical elements are rarely documented in public knowledge bases. When observing a finished product like Apache APISIX, one sees the output but not the thought process, concept abstraction, or the reasoning behind specific architectural choices. AI excels at understanding 'what' needs to be done and 'how' to execute it, but it lacks the ability to judge 'why' a specific approach is superior in a given context, a domain where senior engineers remain indispensable.

To operationalize this insight, API7.ai implemented a controversial policy: engineers are instructed to avoid manual coding whenever possible, delegating these mechanical tasks to AI agents. The primary resistance to this shift came from engineers who had rigidly defined their roles as either front-end or back-end developers. These individuals found the boundary-breaking nature of AI agents uncomfortable, as the tools dissolved the traditional silos that defined their professional identities. A practical example illustrates this shift: previously, a competent front-end developer required deep expertise in aesthetics, performance, SEO, and various frameworks. Now, an individual with minimal front-end coding experience can build a functional dashboard by articulating the evaluation criteria, such as color schemes, CDN usage for resource loading, and mobile compatibility standards. Once these parameters are defined, the AI generates the code, representing a paradigm shift where product development relies on the ability to specify requirements rather than technical implementation skills.

The impact on workflow efficiency is stark. Historically, delivering a user requirement necessitated the collaboration of product managers, architects, and both front-end and back-end developers, a process spanning weeks. Today, colleagues responsible for solutions or even sales teams can use AI agents to make immediate product adjustments and validate them against user needs in under thirty minutes. This closed-loop process has compressed the iteration cycle from weeks to less than half an hour. Those who view this approach as 'crazy' or 'irresponsible' are typically individuals who have limited their own potential by adhering to outdated role definitions or who have not experienced the full scope of AI capabilities. The resistance is often strongest among senior engineers who insist that AI-generated code can only produce trivial products, yet the definition of 'trivial' versus 'production-ready' is increasingly blurred. The critical criterion is not the type of code, but whether the human operator possesses a clear understanding of architecture, code, and testing, along with a sense of reverence for deploying code into production environments.

Human involvement remains essential because even if AI makes 85% or 90% of its decisions correctly, the remaining 10% can significantly degrade overall project quality. No one who cares about technology can produce good code using AI alone without human oversight. Wen Ming established a personal rule: if the logic behind a decision is not understood, the decision should not be made. This principle was applied in the development of AISIX, a new AI gateway built extensively using Claude Code. While AI handled the coding, the design of core concepts, architecture selection, milestone progression, and technical option evaluation were conducted entirely without AI assistance. In this project, AI served as a supplementary tool, controlled through specific directives to perform end-to-end testing, conduct click tests on core dashboard paths, and generate comprehensive documentation. The review process was automated but rigorous: code written by one AI instance was reviewed by a different AI agent, utilizing tools like CodeRabbit and GitHub Copilot for a second layer of scrutiny. Human review was deemed infeasible for the volume of code generated, necessitating an AI-to-AI verification loop.

The true stabilizer of a project, however, is not the elegance of the code or the sophistication of the architecture, but the active usage by a large number of users in a production environment. Only through real-world usage can issues be revealed and addressed through continuous iteration. The revolutionary aspect of AI is its ability to dramatically accelerate this iteration process. For instance, when a user reported a bug at 2 a.m., an AI agent immediately performed a preliminary analysis to identify the affected product, followed by static analysis based on version numbers, context, and error logs. In more than half of these cases, the problem was accurately identified at this initial stage. If immediate resolution was not possible, the AI automatically set up a reproduction environment, ran the corresponding code and plugins through end-to-end tests, and attempted various reproduction methods. Even after identifying the issue, an independent AI agent and environment were used to re-verify the results. Throughout this automated sequence, humans did not intervene in the immediate debugging but focused on optimizing the automated workflow by adjusting prompts, changing models, refining reproduction scenarios, and integrating their own experience. The final decision remained with humans, who could make judgments while AI handled the rapid analysis and environment setup that would be impossible for a human to perform at 2 a.m.

Wen Ming outlines three distinct stages in the evolution of using AI for software development. The first stage involved building a framework using tools like ECC and Oh My OpenCode to identify potential blind spots, creating a sensation of managing a team of multiple AI agents working simultaneously. The second stage marked a shift to abandoning these heavy frameworks and taking back control, driven by the realization that large models are now intelligent enough to search, understand, and complete tasks efficiently without external scaffolding. In this phase, the primary value humans provide is their experience and knowledge, summarized in documents such as agents.md or CLAUDE.md, which serve as references for each task. The third stage represents a transition from 'addictive coding' to 'high-quality decision-making.' As AI accelerates execution, the number of technical decisions required per day surged from two or three to forty or fifty, a volume that exceeds human capacity if not managed correctly.

The term 'addictive coding' describes a period where large tasks were assigned to AI before sleep, with constant iteration and decision-making occurring throughout the day and night. This approach led to a significant decline in productivity due to decision fatigue. The current strategy involves managing five or six development tasks simultaneously, where decisions are made only after understanding the underlying logic and collaborating with AI to reach a consensus. Important architectural choices remain the domain of humans, while machines assist in research, providing background information, filling knowledge gaps, and broadening perspectives. To maintain decision quality, high-quality decisions are concentrated between 9 a.m. and 3 or 4 p.m., after which energy is depleted. The rest of the day is dedicated to reading and exercise to regain energy, with AI usage avoided at night and on weekends to prevent endless iterations that degrade software quality. The expenditure of billions of Tokens is secondary to whether experience is documented and whether high-quality decisions can be made efficiently.