89 sources later, researchers lay out who should approve what

Researchers reviewed 89 papers and found no standard answer to a core question: who approves what an AI agent can reach? Agents can search the web, write files, or call outside services. The paper argues that approving this access has been treated as an afterthought. It belongs at the center of agent design. The researchers identify three decisions a human or system must make. First, what the agent is allowed to reach. Second, under what conditions that access can be granted. Third, who can take it away. Today those decisions fall to whoever sets up the agent. The humans most affected have no standard way to see or adjust them. A designer building an agent product runs into this gap the moment a human hands over a task. The human has no clear view of what the agent can touch. There is no reliable way to narrow that reach before handing work over. The paper treats granting access, limiting its scope, and removing it as three separate things. A human should be able to inspect and adjust each one on their own terms.