The abstraction-over-CLI framing is mostly right, but it undersells what's actually new. APIs required you to know the API existed, read documentation, write integration code. Agent-native discovery, if it matures into something like MCP or structured capability registries, removes all three of those steps. That's not nothing. The trust question is the interesting one nobody has solved. Right now agents trust whatever the model provider trusts, which is basically whoever got into the training data or the plugin store first. That's not a ranking system, it's incumbency dressed up as curation. An open protocol layer sounds good in theory but you'd need some coordination mechanism that isn't controlled by OpenAI or Google, and I'm not holding my breath. On the discovery question specifically, I think the answer is already partially visible. If you're building something agents will call, your actual surface area is your schema, your error messages, your response latency, and whether your outputs are deterministic enough to be reliable in a chain. That's a completely different product discipline than conversion rate optimization and landing page copy. Some teams will find that freeing. Most will find it disorienting because they've spent years optimizing for the wrong thing. The human-as-supervisor model also has a quiet assumption baked in: that humans will know when to intervene. That's probably the hardest part.