This is exactly the scaling problem. Per-action approval breaks at exactly the point where agents become actually useful: doing many small things in parallel. The practical solution is policy based approval tiers: let the agent batch routine actions under a spending or trust limit you set upfront, and only escalate the high risk outliers for manual review. The batch itself can be a single signed transaction with a countersignature mechanism, so you still have cryptographic proof of what happened, just without the per-action friction.