What controls help groups manage AI agent participation?

Having found in Study 1 that group members wanted to control their Slack brainstorming agent "Koala," Houde and colleagues built a second prototype, "Koala II," with functional controls group members could actually operate, then used it to validate and expand a taxonomy of agent-participation controls. The excerpt states the taxonomy "identifies key aspects of control over an agent's interactive behaviors and ways to control them," organized around when, what, and where a conversational AI agent in a group should respond, who can control its behavior, and how those controls are specified and implemented. This differs from the comparison case the authors cite, Gammelgård-Larsen et al., who gave users three fixed configurations (intermittent, continuous, proactive) for a music-recommendation agent; Houde and colleagues instead "examined a set of controls in which participants could create their own configurations" rather than choosing among presets.

The mechanism behind the taxonomy's "when" axis traces back to Koala's underlying operational logic: a user message triggers backend "control logic" that either passes on replying immediately or hands off to further "autonomy control logic," where the LLM generates a candidate response and then scores that same response for how valuable a contribution it would make — the model is asked to judge its own interjection before Koala decides whether to post it. The excerpt also flags a specific failure inside that logic: "the underlying LLM was unreliable in identifying the intended target for a chat utterance," meaning Koala could not reliably tell who a given group message was actually addressed to, a prerequisite for deciding whether it was the intended recipient at all.

The taxonomy sits alongside How can proactive agents avoid feeling intrusive to users? as a second multi-dimensional framework for the same underlying problem, but the two cut the space differently: the Intelligence-Adaptivity-Civility triad describes qualities of the agent's behavior itself, while Houde's when/what/where/who/how axes describe control points a group can act on, with "who" naming control ownership as its own dimension — a question the ICT framing does not pose directly. The targeting failure the excerpt reports also lands squarely inside Why do conversation agents handle turn-taking better than emotion recognition?: addressee selection is exactly the kind of agent-action-modeling task that survey found current systems handle far less reliably than raw language generation, and Koala's difficulty "identifying the intended target" is a concrete instance of that gap surfacing inside a deployed prototype rather than a benchmark.

The excerpt's description of the taxonomy itself is cut off after its introduction, so the full set of control mechanisms validated in Study 2, and any usability findings about which controls group members actually used, are not given here. It is also not established whether a five-axis taxonomy derived from one Slack brainstorming deployment covers the control needs of other group-chat contexts (customer support channels, open public threads, larger groups) where addressee ambiguity and participation norms differ. The implication, at the strength the excerpt supports, is that designing proactive group-chat agents requires treating who decides and how they decide as design variables distinct from when the agent speaks, rather than folding all of initiative behavior into a single proactivity setting.