Keeping S3 off the write path is the design choice that makes the whole pitch click for me: acknowledging a write once two of three availability zones have it on SSD, then batching to object storage in the background, gets durability without paying object-storage latency on every turn. That also explains why separating the agent loop from the sandbox matters, since a long-running turn shouldn't have to keep a heavyweight environment warm just to hold its state. My main question is how that tiering behaves when an actor wakes after a long idle period and its data has fully cooled to S3: is that 138 ms read the typical cost, or does it get worse with larger session transcripts and filesystems?