Comment by monkmartinez

Not really an article, but yeah, I was hoping they went into the underlying mechanism a bit deeper. This paper could be confirmation of what localllamaians have been saying for months; Keep your harness surface small, allow the model to use the harness to build _your workflow_.

I have been doing a LOT of work around this with Qwen3.6 and its been super fun. There are some neat benchmarks that help guide, but nothing beats reading the output... and there is a lot of output to read when trying different quants, etc. Which leads me too...

The other thing I have learned is the "harness" is only as good as the model tuning that goes into it. If your prompt(s) are buggered from the beginning, you are going to have a bad time. The prompt structure and special tokens can be a PITA or really help depending on how much you know.

I don't know how agentic harnesses can work without being optimized for the models running within them. This is the biggest insight into working with agents for me. First thing I have always looked at were the prompts and parameters... everything else is orchestration to me.