AI development as a stack: what to rent, adapt and build yourself (Category AI4SDLC)

Over the past year, I have been actively engaged in the topic of AI development at the level of a large company. On such a scale, you quickly cease to solve only operational problems: you need to define the target picture, agree on what the engineering system should become, and build a strategy for transition to it.

We set goals and strategy as well. But as I moved, I started catching myself with a few less pleasant thoughts.

1а Agent strapping (harness) It's changing with the models so fast that I have a hypothesis that it's half-life in about six months. If it is even close to the truth, investing in your own strapping is almost not capitalized: you run on the spot, like in Alice in the Looking Glass, and to move, you need to run twice as fast.

2New models with open weights increasingly show the effect of joint design of models and iron. Once purchased, the H100 is not a long-term infrastructure strategy. A well-established equipment upgrade channel is needed because model architecture, accelerators and the economy of inferencing are changing together.

3Connecting internal tools requires teamwork and a precise balance between speed and governance. The easiest way to make an agent safe is to limit it so much that it becomes useless. A useful agent requires a much more mature construct: identity, politics, instrument contracts, checks, and a person in the contour of responsibility.

4It is impossible to improve such a system without high-quality telemetry and evals. You need a baseline, playable work episodes, and the ability to consistently check whether a new model, tool, or harness version has really made the system better.

I became interested in how these issues are solved by big players and what to do in such a situation conditional CTO. And the usual choice of build vs buy is no longer enough. It is more useful to divide the stack into three modes: rent, adapt or build your own.

That's it. longrid. In it, I understand AI development not as the sum of a model and a binding, but as a co-evolving system: iron → model → harness → tools → results → tests. At the same time, I check the hypothesis of the harness half-life - the literal period of six months is not confirmed by public data, but the rapid cycle of significant reconfiguration confirms quite well.

#AI #AI4SDLC #Engineering #Architecture #PlatformEngineering #Management