Following that line of reasoning, context length is another very large confounding factor. Longer context lengths improve performance - but also result in enormous increases in KV cache size and memory requirements. We decide to control for this in our paper and focus at the 32K context length for 32B size models, a context length that already pushes the bounds of what can be "deployable" locally.

Still, we evaluate at 64K context length using YARN and are able to outperform CWM's 54% performance (non TTS), which it achieves using 128K context, a substantial increase over what we use. This is also pretty significant because we only ever train at 32K context, but CWM trains for a full 128K.

These all make the "bare LLMs" better suited to be used within the "agent" harness.

I think the more accurate term would be "agentic LLMs" instead of calling them "agents" outright. As to why its the case now, probably just human laziness and colloquialisms.

I give it 3-4 more weeks before we start to hear about the death of agentic frameworks. Pointing GPT5+ at a powershell or C#/Python REPL is looking way more capable than wiring up a bunch of domain-specific tools. A code-based REPL is the ultimate tool. You only need one and you can force the model to always call it (100% chance of picking the right tool). The amount of integration work around Process.Start is approximately 10-15 minutes, even if you don't use AI assistance.

Maybe there's certain problems that it excels at but probably 99% of what I throw it at can be gleaned from the context/nearby code anyway, like you said. Even if I'm using some in-house library (pretty much all of our code), the models are good enough to dig into that library and read the headers if they need to.

Maybe it can help with speed? If it needs to do less research before it can start coding.