The paper studies a persistent Team Agent placed in the collaborative spaces of more than 20 teams at a large technology company. Its stated purpose included helping teams manage information flows and file bugs. Interviews found that, despite broad technical capabilities, the agent struggled with implicit and contextual practices that organize everyday teamwork.
The mechanism is that a shared, proactive agent acts inside social workflows whose rules are often tacit: it must infer relevance, boundaries, authority, and appropriate timing without possessing the relational capital through which people normally negotiate those matters. Proactivity therefore redistributes agency before trust has formed. The authors report that acceptance was mediated by perceived effects on people’s own agency, and that participants treated agent autonomy as a privilege to be progressively earned.
Why now is that enterprise AI is shifting from reactive, single-user tools toward multi-user systems embedded in shared chats, documents, and infrastructure. The implication is that teams need staged delegation, legible controls, and governance alongside model and integration work. A second implication is that success criteria should include whether collaboration norms and human discretion are preserved, not only whether the agent completes tasks. Evidence boundary: this is a qualitative technology-probe study in one large technology company, not a controlled measure of productivity or a general estimate of all workplace-agent outcomes. Watch for whether later deployments publish comparative evidence on which permission models, escalation paths, and team controls allow autonomy to expand without reducing human agency.