Who Started It: Six Capabilities That "Proactivity" Conflates, and the One That Only Ever Points One Way

Two bodies of evidence about conversational agents point in opposite directions and are rarely read together. A four-week study of social chatbot use finds loneliness and social anxiety declining, with self-disclosure to the chatbot the significant predictor in the expected direction; a peer-support agent that opens turns unprompted outperforms an otherwise similar user-initiated one on engagement. Work on proactive agents warns that such systems risk being perceived as intrusive, and reports reduced perceived autonomy and a transfer of control from user to system. The two are not measuring the same actor: the benefits are attributed to something the user does and the costs to something the system does, while a single umbrella term places them in one discussion. That term is currently anchored on initiative-taking as such, which covers several distinct acts. We propose that what is delivered to a deployed system is not a quantity of proactivity but a set of distinct capabilities: to withhold a response, to direct the focus, to end the session, to open a turn unprompted, to introduce a goal the user does not hold, and to carry memory across sessions. We argue they are not orderable on a single axis, name the two candidate orderings we can construct and give our reasons for rejecting them, state the three relations we do accept, and give the observations that would refute the rest. We then examine one capability in detail. Ending the session appears in the current agency literature only in one direction, as the user's ability to abandon a collaboration, filed under outcome control; the system holding the same right is absent, and the axes in use cannot represent it. That capability has already shipped, framed publicly as a question about the model. We report no new empirical results, but a distinction, evidence that it is not currently drawn, and what would show it to be wrong.