As AI assistants juggle multiple people in one chat, who should the rules track: fixed job titles, or who actually knows what right now?

This explores how governance could follow who knows what and who is in charge as those things change during a live interaction, instead of fixing roles in advance. The corpus starts with the gap itself. A scoping review of the OECD, NIST and EU AI Act frameworks finds that they name the affected stakeholders but miss how information, authority and agency move between people within a single conversation with a multi-party assistant Do current AI governance frameworks capture shifting stakeholder roles?. The review's recommendation is to govern at the points where information is handled and actions are taken, not through a list of roles. Agency law and principal-agent economics explain why this matters. Their three classic problems are information asymmetry, discretionary authority and loyalty, and the usual fixes (incentives, monitoring, enforcement) assume someone can watch and interpret the agent's decisions. Fast, opaque AI agents break that assumption Can traditional governance tools actually control AI agents?.

The surprising answer from the corpus is that governance works best when it lives inside the system the agent actually runs on, not in a policy document beside it. In a market simulation, a written anti-collusion constitution did almost nothing. A public, enforced governance graph, with an oracle, a controller and an audit log that can't be altered, cut collusion substantially Can runtime governance reduce LLM market collusion more than written rules?. A long-running persistent agent showed the same pattern. Safeguards written into the memory layer it consulted while working produced 889 recorded governance events over 96 days, because the agent ran into the rules while making decisions Can governance rules embedded in runtime memory actually protect autonomous agents?. So representing authority in real time may mean making it a structure the agent reads and the system enforces, not a description humans keep elsewhere.

The structure has to be wired correctly, though. Older multi-agent organization models treat authority as nested: you can only have authority over a role you can communicate with, and you can only communicate with roles you can represent. But they check this once, when the system is specified, not while it runs How do authority, communication, and acquaintance relate in organizations?. Experiments with RACI accountability tags make the same point. Labeling an agent "accountable" changed outcomes only when the workflow actually routed the deliverable through that agent Does accountability placement actually shape multi-agent outcomes?. A label of authority that the control flow ignores does nothing. Singapore's IMDA framework turns this into concrete design levers: the action space (which tools and permissions an agent has) and the autonomy scope (how much it decides alone) How should organizations govern the risks of agentic AI?. The design community's "autonomy slider", which lets a user set an agent to suggest, draft or execute for each task, is the user-facing version of the same idea Should agent autonomy be a continuous slider or fixed levels?.

The human side matters too. Leonardi's research shows that authority can shift even when nothing formal changes. When people credit an AI with judgment, they feel less in control and less responsible while still holding formal authority. That displacement runs in cycles that can be reversed Does attributing agency to AI actually reduce human sense of control?. In radiology, scheduling and policing, some people gained new authority by managing how agency was attributed, for example as interpreters or curators. Early attribution patterns then locked in which way things went Can humans gain authority by managing how AI agency gets attributed?. A real-time governance record that tracks only formal permissions would miss these felt and social shifts. Meanwhile, BCG's survey of nearly 12,000 employees finds that agent use has more than doubled while half of respondents lack clear governance for human-AI teams How far behind is AI governance compared to agent deployment?.

The corpus has no worked-out schema for logging information and authority changes turn by turn in multi-party settings. That remains an open design problem. What it does offer is a direction: put governance in the runtime, make sure routing enforces accountability, treat autonomy as something you can adjust, and expect authority to shift through attribution as well as through permissions.