# Replies now move to the new server during an update 2026-11-03 · Product > A routine update cut off a long agent reply on 2 November. Running replies are now paused at a safe point and handed to the new server, which finishes them while your chat follows along. On 2 November a long agent reply stopped halfway through, and the chat said "The server restarted while this reply was being written, so it was lost." Nothing had crashed. A routine update had replaced the server, and the old copy shut down while the reply was still being written. Updates here run without downtime. The new version starts next to the old one and takes all new requests. Until now the old copy stopped after ten minutes at the latest. The reply on 2 November was an agent run with many tool steps and needed longer, so the old copy stopped in the middle of it. Now a running reply moves to the new version. The old copy pauses it at the next safe point: right after a tool has finished, or in the middle of the text. Everything up to that point is saved. The new copy picks the reply up within a fraction of a second and continues from there. Your open chat reconnects on its own and shows the rest. A tool never runs twice. If an update hits a tool that takes longer than 20 seconds, the tool is stopped and the model is told that its result is unknown, so it checks the state before trying again. If the text was cut off, the new copy writes that step again and the half-finished text is replaced, so it doesn't appear twice. Every model call is charged once. If a cut-off call reports no usage, we book an estimate for it instead of nothing. A first fix on 3 November let the old copy wait up to 32 minutes instead. It held up the next update and was rolled back after about half an hour. Nothing to do on your side.