The portability angle is interesting. We measured coding-agent sessions and found ~94.5% of the accumulated context was reused across turns.
Two questions on the translation side: what gets dropped or normalized when a session crosses agents (reasoning blocks, tool-call IDs, provider-specific message types)? And is the output serialized deterministically, so the receiving agent's cache can warm up and stay warm after the move?
Two questions on the translation side: what gets dropped or normalized when a session crosses agents (reasoning blocks, tool-call IDs, provider-specific message types)? And is the output serialized deterministically, so the receiving agent's cache can warm up and stay warm after the move?