Biggest mistake with AI agents right now:

giving them tools and only checking after the damage is done.

Zelta flips it. Approval happens before execution. Token is consumed on verify. Replay and tampering get rejected.