Behind Coze's Shutdown: The End of the AI Workflow Era and Big Tech's Path Dependency

ByteDance's AI workflow platform Coze is about to go offline after burning through 1 billion yuan, a piece of news that has sent shockwaves through the AI community. As one of the hottest AI products of its time, Coze's demise is not just the failure of a single product—it reflects the collective predicament of the AI workflow form and the deep differences between Chinese and American AI development paths.

Why Coze Fell from Grace

Coze's rise stemmed from an era when large model capabilities were still immature. At the time, models couldn't reliably complete long tasks, so workflows broke tasks into multiple steps, hard-coded the first few steps, and only called the large model at key nodes—thereby reducing hallucinations and ensuring output quality. This engineering approach quickly made Coze a star tool for AI application development and spawned a wave of similar products such as Yingdao RPA, Dify, and N8N.

But the problem is that this path is essentially about building AI products with traditional internet thinking—using fixed processes to compensate for insufficient model capabilities. As large models iterate far faster than expected, especially with the recent breakthroughs in AI coding capabilities, models can now autonomously plan and execute multi-step tasks. At this point, the logic of pre-orchestrated workflows appears clumsy and inefficient, making Coze's decline inevitable.

AI Workflow is a development paradigm that breaks complex tasks into multiple steps and orchestrates their execution order through visual connections or flowcharts. On platforms like Coze, users can split a task (such as "generate an industry report") into nodes for retrieval, summarization, writing, and review, where each node can either hard-code fixed logic or call a large model. The advantage of this approach is strong controllability and stable output, especially suited to a stage where model capabilities are insufficient and hallucinations are common. But it essentially uses human-preset processes to constrain the model, assuming the model cannot autonomously complete multi-step reasoning. As large models' autonomous planning capabilities grow, the many "manually orchestrated" nodes in workflows become efficiency bottlenecks—tasks the model could complete in one step are split into multiple steps with human rules added, increasing latency and maintenance costs.

Workflows Are a Transitional Form; AGI Is the Endgame

The core contradiction of workflow products is this: they assume large model capabilities are limited and require humans to pre-design processes. But the ultimate form of AGI requires models to autonomously complete all reasoning and logic rather than relying on hard-coding. When AI coding tools can autonomously untangle complex problems and plan routes, the value of workflows is greatly diminished.

The blogger candidly admits in the video that he once built a workflow platform based on open-source Coze, only to realize a few months later that the direction was wrong. He judged that the iteration speed of large models had been underestimated—he originally thought it would take three to five years to reach the level of autonomously executing multi-step tasks, but progress came much faster. Today, simply letting AI coding autonomously plan routes (i.e., Harness engineering) has become a more efficient solution, while the workflow-style step-by-step wiring and prescribed outputs seem "rather silly."

Harness engineering is a practice that has emerged in the AI coding field in recent years, referring to building an "execution framework" (harness) for AI agents in which the model autonomously plans, calls tools, writes and runs code, rather than having humans prescribe every step in advance. Unlike workflows, Harness does not preset a task decomposition path but hands planning authority to the model itself, with humans only providing goals, tool interfaces, and verification mechanisms. This model depends on the model having strong code generation and self-correction capabilities, which is why it only became feasible after breakthroughs in large model coding abilities. Its core value is that the model can dynamically handle unknown tasks, unlike workflows that can only handle pre-imagined scenarios. The article's phrase "simply letting AI coding autonomously plan routes" is precisely an expression of the Harness approach.

Diverging AI Paths Between China and the US: Commercialization Anxiety vs. Ultimate Questions

Coze's fate highlights the path dependency differences between Chinese and American AI development. US giants like OpenAI and Anthropic (Claude) aimed at the ultimate questions from day one and were not in a hurry to launch intermediate transitional products, so they wouldn't invest resources in forms like workflows. Chinese giants and startups, by contrast, face pressure for capital returns and must commercialize as quickly as possible, which gave rise to products like Coze that stabilize model capabilities through engineering means and monetize them.

This path choice leads to a dispersion of effort: continuously refining transitional products means being unable to focus on solving the ultimate questions. Over time, the technology gap gradually widens. The blogger points out that among Chinese giants, Tencent's Work Body essentially has no intelligence—it's just an application-layer product calling others' models; ByteDance's Doubao and Feishu are also concentrated at the application layer, even using AI to make short dramas (Hongguo Short Drama) for monetization. The only one truly aiming at AGI is DeepSeek^Ad, whose open-source Harness Agent is not for profit but for collecting coding data to train models—coding data is the highest-quality data.

AGI (Artificial General Intelligence) refers to AI systems that can match or surpass human levels across a wide range of tasks, rather than excelling in only one specific domain. In the context of this article, AGI means models can autonomously complete reasoning, planning, tool calling, and multi-step execution without humans pre-orchestrating processes. The opposite is "application-layer products"—building interfaces, processes, and scenario wrappers on top of existing model capabilities, without pursuing fundamental breakthroughs in model capabilities. When US giants aim at "ultimate questions" from day one, it means investing resources in raising the capability ceiling of the models themselves; Chinese giants, due to capital return pressure, are more inclined to quickly produce monetizable application-layer products. The divergence between these two paths determines the fate of transitional forms like workflows in different markets.

Application-Layer Involution and the Predicament of the Startup Ecosystem

Chinese giants clustering at the application layer has led to a distorted AI startup ecosystem. The application layer is essentially a traffic business—whoever has traffic can capture users, leaving small startup teams almost no opportunity. The blogger believes that giants should work on foundational models and tools like Codex and Claude and leave the application layer to small teams, but Chinese giants do exactly the opposite.

The result is that China's only purely AI company is DeepSeek. The blogger even asserts: "If DeepSeek can make it, China can win; if DeepSeek can't make it, China's AI is finished." Meanwhile, giants pour massive resources into monetization businesses like short dramas—Hongguo Short Drama has 160 million daily active users and 120 minutes of daily use per person, which the blogger sees as a waste of precious time.

Coding data is regarded as one of the highest-quality data types for training large models because it contains rigorous logical structures, clear inputs and outputs, and verifiable correctness. Code naturally carries a "right or wrong" signal, and models can verify their own outputs by running tests, making coding data extremely efficient at improving model reasoning capabilities. DeepSeek open-sourcing the Harness Agent and letting developers use it for free is essentially collecting coding interaction data from real scenarios to feed back into model training. This is completely different from the logic of simply monetizing at the application layer: the former accumulates model capability assets, while the latter consumes traffic dividends. Understanding this makes it clear why the article considers DeepSeek's path unique among Chinese AI companies.

Conclusion: The Dusk of Workflows, the Dawn of AGI

Coze's shutdown is not an isolated event; it marks the end of the era of AI transitional products represented by workflows. When model capabilities are strong enough, scaffolding like workflows will inevitably be abandoned. For practitioners, recognizing technology trends and avoiding over-investment in transitional forms may be more important than chasing short-term commercialization. And whether China's AI can win the AGI race depends on whether more companies are willing to solve the ultimate questions rather than becoming addicted to traffic monetization.