that's right Andrey, I'm an architect.
Base on my expreince: The most difficult part of building software applications is orchestrating various services: account systems, payment gateways, databases, security, domains...
CI/CD pipelines, network configurations, and so on.
These constitute a massive system that needs to be managed completely outside of the code itself.
It's not merely about the core code of the Agent—AI is already capable of handling that part.
What actually falls outside the reach of large models are precisely those infrastructural elements just mentioned.
A significant portion of this relies on the years of experience and soft skills accumulated by architects and engineers, which simply cannot be translated into explicit code.
Even for the exact same business goal, the holistic and localized approaches will vary entirely depending on the organizational context and the target user demographics.
This is exactly why the concept of "Harness Engineering" was brought forward in 2026: --> It emphasizes a stronger focus on engineering capabilities—essentially, the "environment" surrounding the Agent.
Hopefully, we will see substantial advancements in Harness Engineering this year.