What should require human approval?

The biggest risk with an AI agent may not be what it thinks. It may be what it’s allowed to do.

A model can be impressive at planning, but once it can spend money, call paid APIs, or act across services, a bad decision stops being just a bad answer. It can become a real transaction.

What caught my attention in Kite’s design is the separation between the user, the agent, and the session. The user remains the root authority; the agent receives delegated permissions; and a session can be limited to a specific task or period. That creates a clearer boundary between asking an AI to help and giving it the power to act.

Kite’s approach also puts programmable constraints at the infrastructure level. Spending limits and permitted actions are meant to be enforced through cryptographic authorization and smart-contract logic, not merely written into a prompt and hoped for the best.

To me, that’s a more useful way to think about AI safety in financial workflows. We probably won’t eliminate every model error. But we can design systems where one error doesn’t automatically become unlimited authority.

The question I’m left with is this: if an agent has a clear identity, limited permissions, and an auditable trail, what should still require human approval? I’d draw the line differently for a small API payment than for moving funds between accounts.

That permission boundary may be just as important as the intelligence behind the agent.

#KiteBBW $KITE @KITE AI