Qwen3.6-35B-A3B spirals in tool loops

This reads like a stability problem more than a capability problem: if a model cannot stay on task inside an edit loop, raw benchmark gains do not help much in real coding work.

• –The report is anecdotal, but it is the kind of anecdote agentic model users care about: control-loop discipline matters as much as output quality
• –A model that keeps “thinking in circles” can be worse than a less capable one that makes steady edits and stops cleanly
• –Local runners like `ik_llama` may expose tool-use and stop-condition issues that are less obvious in hosted demos
• –If Qwen3.6 is meant to be better at frontend and agentic coding, it needs to be judged on loop stability, not just reasoning claims
• –The useful benchmark here is not “did it produce a nicer button” but “did it converge without self-reinforcing retries”