Multi-agent approaches (to side effects) — flags as exceeding its own frame → Shutdown / off-switch problem

Multi-agent approaches reframes side-effect avoidance as a matter of "understand[ing] all the other agents (including humans) and mak[ing] sure our actions don't harm their interests" (concrete-problems, §"Multi-Agent Approaches:", p. 6), and shutdown is introduced as one application of that reframe via CIRL. But the paper immediately hedges on how well the reframe contains it, calling the shutdown issue "an interesting problem in its own right" and pointing to dedicated treatments elsewhere, Orseau and Armstrong's safely-interruptible-agents work and Hadfield-Menell's off-switch paper. That hedge is a tell: shutdown is not fully absorbed by the general "protect other agents' interests" framing the way the rest of multi-agent approaches is, because refusing shutdown is a specific, structural failure, resisting correction itself, rather than one more ordinary externality.