I think everyone needs to fight to keep standards around AI tooling open, even if proprietary models are inevitable given the investment they take. AGENTS.md is a good example. Every Claude integration (Github CI etc), every shared artifact your boss comes to rely on entrenches us further. I think where possible all AI outputs need to be persisted in your own systems For example rather than have Claude create an artifact, have it create an HTML app / report in a repository you control. Make your own tools if you have to. I'm working on building a Formal Verification system for reviewing agent output into my notes / project management tool https://app.etherpk.com based on local markdown graphs for example.
It requires a LOT more effort, but all of my tooling is non-agentic and fully model independent. If necessary, I can swap it all out to local or private cloud services.
It results in a far more involved development process, which helps keep me in the loop but also has a big speed impact.
Not handing it all off to agents can trigger some FOMO. I'm not sure how long I can keep it up and still compete.