As AI agents use more tools and APIs, where do you think accountability should start?

When an agent can make decisions and execute actions across multiple systems, should every step be traceable, or is the final outcome enough?

I am curious how people see this.