[avereveard]

"Minimal infrastructure, max LLM freedom" works great for personal automation. The same shape under enterprise security review collapses on the question of what cannot happen, which is exactly what the prompts-and-vibes school doesn't have a structural answer for. Direct CDP hands the model the keys; the harness around it is what should decide which doors the keys open.

Most agent stacks at AI startups have that layer as llm driven glue rather than an owned surface, and it shows up as a re-architecture cost on every model release. model should be replaceable, the integrations and guardrails specific to the customer's environment should not.