LangSmith Unveils Engine To Power Enterprise AI Flywheel

The challenge is that the frontier labs are constantly iterating on both their models and their harnesses, and those models are trained and optimized against their own harnesses. In contrast, when a person or small team typically gets access to a more generic harness. Deep agents I'd argue is probably the best general-purpose option available, but it's difficult to keep pace with the labs in terms of reliability and consistency.

Using another company's model and harness is an easy choice because you get support and a mature ecosystem, but there are definite tradeoffs.

What I'm curious about is whether your approach is to build around the existing harness and model, letting your solution adapt to them, or whether you modify the harness and model to fit your own requirements.

Many people would instinctively choose the latter, but for a small team there are significant infrastructure and operational challenges that come with maintaining a customized stack.