Andrej Karpathy mentioned that the most difficult part of building an intelligent application is assembling various services, account systems, payment systems, databases, security, and domain names.
.
CICD, network settings, etc.
are all huge systems that need to be processed outside of the code.
It is not just the core code of the Agent itself.
This part of AI can already do it.
Often what is beyond the capabilities of large models is exactly what I just mentioned.
Many of these are architects and engineers who have accumulated many years of experience and soft skills that cannot be converted into clear code.
At the same time For the same business goal, in different organizational environments and different user types, the overall and partial methods adopted are different.
This is why the Harness project was proposed in 2026 --> Pay more attention to engineering capabilities, that is, the "environment" around the Agent.