Inside China’s Tech Companies, a Battle Over the Soul of Software Engineering
Walk into any of China’s major technology companies today, and you will find two kinds of software engineers living in the same building, often on the same floor. The first kind uses AI to run hundreds of parallel tasks simultaneously, such as writing code, reviewing pull requests, generating test cases, and debugging production issues, all with minimal human intervention. The second kind refuses to let AI touch their work at all, insisting that the quality, security, and intellectual integrity of their code depends on human judgment that no model can replicate.
A new investigation by TMT Post documents this divide in striking detail, drawing on interviews with engineers at Alibaba, ByteDance, Tencent, Meituan, and several AI startups. The piece describes a profession in the middle of a genuine identity crisis, one that is playing out not just in China but across the global software industry, with China’s particular intensity and speed.
The 600-Agent Engineer
The most extreme practitioners of AI-assisted development described in the TMT Post investigation are running what they call “agent farms,” networks of AI agents, each handling a specific subtask, coordinated by a human engineer who acts more as a project manager than a coder. Liu Hao, a senior backend engineer at a major internet company, described stuffing more than 600 parallel tasks into his own agent in the month after the Chinese New Year, spanning over 20 vertical scenarios and accumulating more than 100 skills. His monthly token bill exceeds 10,000 yuan and is still rising.
The productivity transformation he describes is dramatic. A task that once took seven days of development can now be completed in two hours; a two-week requirement can be delivered in two days. Liu Hao now spends more than 16 hours a day in front of his computer, not writing code, but orchestrating agents, stating requirements, reviewing results, and feeding new tasks into the system. A tech VP at a large company reportedly tried to recruit him to lead an internal AI transformation team; Liu Hao’s agent-managed calendar was already too full to even schedule a meeting.
This pattern is consistent with broader data on China’s AI labor market, which shows that demand for senior AI engineers is growing while demand for junior roles is stagnating. The AI agent revolution is not eliminating software engineering as a profession; it is bifurcating it, creating a small class of highly productive, highly compensated “conductor” engineers and a shrinking market for traditional junior developers.
The Refusers
At the other end of the spectrum are engineers who have deliberately chosen to keep AI out of their workflow. Their objections are varied: some cite security concerns, noting that AI-generated code can introduce subtle vulnerabilities that are difficult to detect in code review. Others raise intellectual property issues, arguing that models trained on public code repositories may reproduce copyrighted material. Still others simply distrust the output, having encountered enough AI-generated bugs to conclude that the productivity gains are not worth the debugging overhead.
The TMT Post investigation found that the “refusers” are disproportionately concentrated in industries with high security requirements, such as finance, healthcare, and defense-adjacent technology, where the consequences of a code error are severe. Several engineering managers at financial technology companies described explicit policies prohibiting the use of AI coding tools on production systems, citing both regulatory requirements and risk management concerns.
The Management Challenge
The divide between AI adopters and refusers is creating significant management challenges. Teams that mix the two approaches struggle with code quality consistency, review standards, and knowledge transfer. Engineers who use AI extensively often produce code that is difficult for non-AI users to understand or maintain, creating a form of technical debt that is not immediately visible but compounds over time.
Several engineering directors interviewed by TMT Post described the challenge of setting team-wide policies in an environment where the technology is evolving faster than any policy can keep pace with. The consensus, to the extent there is one, is that AI tools should be permitted but that human review of AI-generated code should be mandatory, a standard that is easier to state than to enforce.
The divide documented by TMT Post is not unique to China, but it is playing out with particular intensity there, given the speed of AI adoption and the competitive pressure that Chinese tech companies face. As AI commercial applications spread across China’s economy, the question of how to integrate AI into software development workflows and how to manage the human consequences of that integration will only become more pressing.
What Comes Next
The TMT Post investigation does not offer a resolution to the divide it documents, because there is none yet. The technology is moving faster than organizational norms can adapt, and the right answer, if there is one, will likely vary by industry, by company culture, and by the specific nature of the work being done. What is clear is that the divide is not going away. The engineers who have embraced AI agents are not going back, and the engineers who have refused are not going to be persuaded by productivity statistics alone.
The more consequential question may be what happens to the junior developers who are being displaced by AI agents before they have had the chance to develop the senior-level judgment that makes human engineers irreplaceable. If the entry-level pipeline dries up, the supply of experienced engineers will eventually shrink, a dynamic that could create a talent crisis of a different kind in five to ten years. China’s AI talent data already shows stagnating demand for junior roles alongside surging demand for senior ones. The long-term implications of that shift for the profession are only beginning to be understood.