Qwen 3.6 27B fails at agentic work despite strong single-prompt performance

A user reports that Qwen 3.6 27B performs well on single prompts but fails significantly during agentic tasks. While the model generates impressive HTML and longer content than the Qwen 3.5 series, it makes continuous errors and ignores directions in multi-turn scenarios.

• The model produces high-quality demo HTML pages and generates much longer content than Qwen 3.5 models.
• In agentic workloads, the model falls apart, making mistakes every four turns or so.
• The user reverted to using Qwen 3.5 122B after multiple failed attempts to stabilize Qwen 3.6 27B.

The experience suggests that while Qwen 3.6 27B is capable for isolated tasks, it lacks the reliability required for complex, multi-step agentic workflows.