Can we make agents more reliable with better deterministic tooling?