I was reading a whitepaper about DeepSeek’s “ah ha” moment a few weeks ago. The authors walked through why chain of thought usually does not work the way people assume.
The key concept was that it is not the chain of thought that gets you to a useful place.
In order for the process to work, the model has to be capable of self reflection. Able to think about its own thinking. For the last three years, I have been calling this metric “recursive depth,” which is the space a model has to evaluate and re-evaluate its own reasoning.
Model size matters less than training methodology and the way the model handles uncertainty. Chain of thought can help, but often because it stabilizes the model’s internal certainty levels instead of magically unlocking better reasoning.
If any given model has the recursive depth to manage the process, more time has a measurable effect. If not, you get slop in your context window.
The harness can own the process
But here's where the magic happens.
Data scientists like to build everything into monolithic harnesses inside the weights themselves. There are reasons and situations where this might be helpful. Although... I mean, in strict technical engineering terms, it's not a requirement.
You do not have to build this capability into the model in order to use this particular set of insights.
If you have a model that naturally reflects and simply needs more reasoning time to achieve a better output, you can put that behavior into the agent harness or the inference system.
This is process... not religion.
Spend reasoning where it matters
The trick is conserving your resources.
Not every response needs deep reasoning.
It still surprises me how naive some of these processes are. You don't need, what was it, 20 minutes of reasoning for low stakes inquiries.
Clearly, the correct approach is to wait until the model is about to hallucinate or improvise, when its internal certainty drops.
That's the moment when you trigger a deeper reasoning and/or research pass.
You measure the risk of a wrong answer, and only when it matters, you give the model more room to contemplate a solution.
And that changes the entire workflow.
If at the end of the process, the agent is still unsure... you instruct it to do the unthinkable. Have it ask for clarification, more information, or question why the user is asking for this specific thing.
In the world, content was king.
But with reasoning systems, this idea is a grave mistake.
In the universe of reasoning systems, context is king.
Ponder... or perish.
The thing I find fascinating about it is the way it seems to work against scaling laws. You see, this methodology can increase your reasoning density per gigabyte. However, it doesn't always scale linearly.
If everything goes well, I will have a demo this week.
And you will be able to try it yourself.