I've spent the last couple of years deploying enterprise AI, and the fear of 'vendor lock-in' is misplaced. The Salesforce playbook doesn't work here because models are stateless APIs and intelligence costs are collapsing (265x in 3 years). The real constraint isn't compute — it's energy. When Microsoft is restarting Three Mile Island and Google is commissioning nuclear reactors, that tells you where the actual bottleneck is. The lock-in risk isn't the model you choose; it's the architectural debt you build around it.
Author here. The key distinction I'm drawing is between sub-agents (ephemeral, spun up per task, no memory) and what I'm calling mesh agents — persistent, with their own accumulated context and boundaries set permanently by the human, not by whatever orchestrator calls them.
Most multi-agent frameworks treat agents as function calls. This is about agents as peers with standing.
The thing I didn't say explicitly in the post: most people's instinct is to connect the agent to everything first and worry about trust later. I did that too. The discomfort was the signal.
Had an email exchange with a solo developer who builds nothing with AI — his app is beautiful. Made me think about where craft behind building software goes, not whether it survives the changes that AI is bringing.