> I've been wondering, who builds tools now if AI are the primary users?

AI already builds tools for itself as part of solving a problem. I expect these to get more sophisticated with time.

> Will AI agents build and publish tools for the greater good? Or will vendors hold back AI generated composable modules as a competitive advantage.

Oh they definitely will - it'll be their last ditch attempt at staying relevant, after they get over denial phase and realize that products are already being subsumed by general AI systems, and reduced to bag of tools agents can use.

But then, for any given bag of tools, how hard will it be for an agent to build a clean-room version of it, instead of paying the vendor? This is the end of software companies as we know them.

> Similarly for languages, I wonder if or when AI generates machine code and skips all our higher order languages we've built for human needs.

That's currently too big a target to hit - the weight of all pretraining, and existing infrastructure, will keep programming languages alive for a while (human languages, that's another topic - there's not as much baggage for agent-to-agent communication).

It will take time. We've only started the co-evolution process very recently: we built agent-optimized wrappers around CLI tools and APIs, and around now that should've percolated into training data for the current or next generation of models. I imagine over time, the tools and agent preferences will co-evolve towards an equilibrium.