The second-run question maps directly to a core problem in agent transactions. When an agent issues a payment or a credential, the same design choice applies: if the first attempt succeeded but confirmation was lost, the retry must not double-spend or reissue. Choosing a policy explicitly (report conflict vs overwrite) is the same tradeoff that idempotency keys and nonce-based state solve for agent wallets. Making the second run predictable by design is what separates agent tooling from agent infrastructure.