Coze shutdown hint sparks workflow vs. agent debate

On November 2, 2026, an employee working on Coze at ByteDance posted that Coze may be shutting down. The employee said his profile still listed him as a "Coze Workflow 引擎开发者" and suggested that by next year few people would know what Coze referred to. He followed up with a few more posts describing what outsiders cannot see. The second was the most striking: he said the drag-and-drop nodes the team worked so hard to build were, in his view, compensating for weak model capability and were ultimately a transitional pattern inherited from low-code tooling, a middle-stage product rather than an end point. The post then turns to a practical concern for AI builders: whether to draw workflows or hand decisions to the model. The author notes that when someone who built the engine describes his own approach as a middle-stage product, it unsettles teams that chose workflows, even though workflows may still be the best fit for tasks needing reliability, accuracy, and reproducibility. To illustrate the shift, the author describes swapping the underlying model to Claude Opus 5.5, after which a set of carefully tuned higher-level skill wrappers no longer worked as well; the model reportedly said those skills limited its performance and that it had a better approach. The post frames this as the same phenomenon as the Coze comment, at a different scale: once a model can plan, scheduling control shifts away from a fixed canvas. As of November 7, 2026, the only evidence reported is the employee's implication and the subsequent commentary; there is no confirmation that Coze has shut down.