Reasoning can be fun.
Watching your agents reasoning is helpful and funny
September 04, 2026
I definitely recommend watching the raw reasoning stream from an agent while you're doing any sort of novel or important work. It's important to understand what an agent is doing and their confidence with an approach. I had tasked a Qwen 3.8 Flash Next agent with building an isolated setup for Hermes, so I could micro control network access even beyond what the system has access to. Additionally you should never curl | bash install things, a simple change to an installer script will essentially hand access to your system over to a bad actor. Having an LLM build your own scoped installer is a good idea.
Pretty quickly into this task Qwen via pi.dev got itself into an amusing loop.
Genuinely gave me a little chuckle. It had reasoned its way out of it with a "I'm going in circles." "Fine, don't nitpick" before I could even read it and had moved on, completing the task rapidly. Well done.
All reasoning models can loop a bit, I suspect this is exacerbated by it being an NVFP4 quant. My guess is there is enough logit flattening such that the model to grabbed an essentially equivalent but different token representing the same URL. It is exciting that we've come so far that even tiny quants like NVFP4 remain usable despite hiccups like this happening under the hood.