The constraints are human now
What I’ve been thinking about this week:
- Models are getting so good, but adoption is way slower. It’s impossible to stay on top of progress even with AI, so there’s a bigger and bigger gap between what you can do and what is being done.
- As agents get better, they are also using more tokens (currently). Sometimes an offshore human is probably cheaper. The financial and budget layer for managing them is going to be increasingly important.
- Consumer agents are going to be massively subsidised as a distribution war. We all win. I’m running cheap fast experiments with those free tokens, then moving winning ideas to my paid models (Claude and Codex mainly).
- Open source is everything. When the agents can read the code they can modify it just the way you want. This is a superpower for those willing to embrace it. I’m testing moving to Linux.
- It’s clear companies need to be accountable for their agents. Governed authority will become a product category — companies need systems that show what an agent knew, how it knew it, what authority it had, what it did, and which human approved these boundary conditions.
Since Astra, our AI capability is no longer the scarce input. Now the constraints are much more messy and human: organizational adoption, trusted context, action authority, energy, permission to build, and economic controls.
I’m launching something that ties these ideas together soon — can’t wait to share it with you.