Coze's Sunset: Workflows Weren't a Mistake, Just a Stopgap
The Coze shutdown and the Opus 5.5 skill failure show that bolted-on tools for compensating model limits have a short shelf life. Teams that build workflow scaffolding around a specific model's weaknesses risk paying for it twice: once in maintenance, and again when a model upgrade renders the layer counterproductive.
A developer who built Coze's workflow engine at ByteDance recently posted that his profile still lists that title, but next year nobody will recognize it. His second point was blunt: the drag-and-drop nodes he had worked on were a patch for the model's inability to plan, an intermediate product carried over from low-code. That struck a nerve for anyone who chose visual workflows over letting the model decide.
The doubt quickly hits a counterexample. The same developer switched to Claude Opus 5.5 and found that a set of orchestration skills, meta-skills that managed lower-level tools, became a liability. The new model said it had better solutions than those hand-coded skills. It's the same story as Coze at a different scale: when models were weak, tooling compensated; when they strengthen, that tooling becomes a constraint. But the question 'was workflow wrong' assumes new is always better. Workflows optimize reliability, accuracy, and reproducibility, which are not the same thing as flexibility. Video production is a case in point: changing one frame on delivery must not surprise the client, so an agent that re-plans each run is unacceptable.
The article splits agents into SOP and autonomous. Workflows fit processes with a known standard answer; self-planning fits coding, debugging, and open research where each step depends on the last. Coze's low barrier also made it a cheap on-ramp for beginners and small businesses, and even if the product dies, the skill of decomposing work into steps persists. The deeper takeaway is that tooling should span the model capability curve: good tools don't constrain a stronger model and still lift a weaker one. The Opus 5.5 experience is a reminder that the same skills can flip from an asset to an obstruction as the model underneath changes.
The 'intermediate product' admission is a useful forecasting tool: if a tool exists mainly to patch a model's current blind spots, expect it to be discarded once those spots close. The Opus 5.5 case proves that even meta-skills can turn negative after a model upgrade, so investments should favor abstractions that don't encode assumptions about model planning ability.
The author's distinction between reproducibility and adaptivity is the practical crux. It sidesteps the hype cycle of 'agents replace workflows' and instead forces a task-based analysis, which survives model generations.
Coze's low barrier may be its most durable legacy: it taught thousands of beginners the visual decomposition of work into steps, and that mental model will persist even if the product vanishes.
May I ask, after a model upgrade, the skill that manages skills doesn't work well. How to understand this? Why would a structured text defining a workflow become ineffective due to model upgrades?