Giga Conf: What Should Happen to a Harness After a New Model? (#AI4SDLC)

Today, October 16, I’m speaking at Giga Conf with a talk titled “Harness: the past, present, and future of the layer between the model and the world.” You can join the livestream at 16:00. I’ve talked a lot about agent harnesses lately: tools, memory, context, checks, and all the everyday engineering that helps a model carry a task through to a result. In this talk, I want to put the whole picture together and tackle an uncomfortable question: how much of what we’ve built is still necessary after the next model comes out?

A harness gradually accumulates instructions. The model forgot to check its work here, so we added a rule. It lost state there, so we came up with another step. Then the model was upgraded, but all those workarounds stayed. We end up carrying around a museum of previous generations’ limitations :)

In the talk, I’ll trace the path from early agent loops to today’s Claude Code, Codex, and Cursor. We’ll look at what they’re built from and where execution state is stored. Then we’ll explore three ways of viewing the same harness: as an all-in-one product, a standalone agent, or a component of your own system.

The slides are already on my website, but today’s talk is limited to 20–30 minutes. If you’d like to hear the director’s cut, give this post 100 🔥 reactions and I’ll do a livestream with an extended version of the talk.

#AI4SDLC #Agents #Engineering #Architecture #Conference