A long-running agent session outlives the network it started on. Four separate mechanisms keep conversations alive across the things that would otherwise end them.
While the link is down
Frames buffer, within bounds. Transcript events are re-sent from the last sequence number the console acknowledged, rather than queued without limit — so a long outage degrades instead of exhausting memory.
When a machine stops
A host that is shut down hands its conversations back rather than ending them, and the fleet picks them up when the machine returns. The policy is a setting: resume everything automatically, resume only the one that was mid-turn, or ask.
Resumption is staggered, so twenty machines rejoining after a network event is not a stampede.
When a machine simply vanishes
A host that disappears without closing its connection is swept by a reaper, so the rail cannot go on claiming a dead machine is fine.
When it comes back
A reconnecting machine declares what it is actually holding. Anything the console believed was running there but is not gets ended, with a line in the transcript saying so — rather than a tab that quietly lies.
One honest limit
A host restart means the agent processes themselves are gone. The conversation is picked up again by its session id; the running process is not resumed.