And in a typical enterprise environment dealing with documents, images, and broader business reasoning skills matter. Agents are not just for code and text :)

In my experience it is almost never productively thinking past that point, just spinning in circles. I also reinject all of the thinking. And there are a few tells that it is getting stuck. I have an optional mode that takes the last few turns and tool calls and shoots it off to DSV4 with a prompt to basically understand where it is at and inject better thinking and or planning. It almost always gets it over relatively difficult humps, but some of the time I don't want things going remote. It might end up with 10-30 cents of DSV4 calls over a hours and the quality improvement is remarkable.

The other thing is I trick it into thinking a web_search tool is a web search but it really just asks DSV4 the prompt. DSV4 is a cheap filter to help prevent prompt injection lol. You can give it other models but DSV4 is my cheap-mode default.

edit: oh! My final 35B A3B tip -- use an extremely simple harness. Pi is good. Pi's default tools almost exactly match what Qwen says they tested the model with (likely meaning that tool set is also what they trained it with or something similar). So, in my experience bigger harnesses don't have a noticeable improve ment on tasks.

Personal: Local models. Agent harnesses. Track cars, turning wrenches.

Had agents build recently:

- Photo catalog app with all 100K photos in my library described by a local 9B parameter model (it is absolutely amazing and my favorite personal project in terms of genuien impact)

- Generic RAG layer (vector embedding + FTS in SQLite with RRF and document ingestiation layers for 1-2GB knowledge bases to power agent context for stuff I am interested in, neuroscience and philosophy is where I spend time on this right now, and basically every AI research paper publicly available to help agents tinker with local models). This is especially useful for local models, but it even helps frontier models because they can freely churn through without a bunch of web calls and such when you are on a topic thread or research.

- Qwen 3.6 35B A3B clean-up proxy, achieving a clean thinking + tool calling A3B that I can leave running for 24+ hours. This particular model gets stuck in long reasoning chains and needs encouragement to get on with it, especially on context tasks. It actually reasons well but then starts second guessing and loops. Built a deterministic framework to inspect its thinking and make it wrap its thinking up if it gets stuck and also to trigger calls to external models and inject solutions to problems when it gets stuck. The agent sees the solution injected from a remote model like it is its own thinking, kind of a subconscious it has no idea about. Metacognition in models, it is wild.

DSV4->GLM->Kimi is the "phone a friend" external brain pathing I built. As far as I know no one has shared all of the bugs I have 'fixed' in my project and gotten local Qwen A3B this reliable.

One of the reasons is, with good specs and design, A3B is just so fast. It isn't as smart as the 27B model, but it is close enough it can usually figure it out with the right tools.