Philipp Schmid of Google DeepMind describes an AI coworker as an agent with a model, a computing environment, instructions, tools, credentials, rules and a schedule. He demonstrates a managed agent that uses files and skills to perform a portfolio audit.
Philipp Schmid explains how a network proxy can substitute credentials for placeholder values on permitted requests, and how hooks can require approval for selected actionsAn agent permission boundary defines and enforces which resources and actions an automated agent is allowed to use.. A scheduled demonstration retrieves financial information and updates a spreadsheetWorkflow automation uses software or AI to complete a connected sequence of routine steps with less manual effort.; his security claims describe the design presented in the talk.
Watch on YouTube
