Qwen 3.6 27B absolutely fails at agentic work

I have been running Qwen 3.5 122B at 4 bit for quite a while, and have started running it at 5 bit recently now that Llama.cpp has comparable performance to VLLM. I have also tried, several times, to use Qwen 3.6 27B at 8 bit & 16 bit, as numerous people have claimed that 27B is better than 122B. And it is, on single prompts. It will output very impressive demo HTML pages. It has the ability to generate much longer content than any of the 3.5 series models. However, on agentic work, it absolutely falls apar