A multi-agent workflow is a process in which two or more AI agents hand off data, decisions, or actions to one another to complete a task. For example, one agent may gather information, another may evaluate it, and a third may use a tool or service based on the result.
These handoffs can make it difficult to determine which agent initiated an action, what information influenced it, or whether anyone approved the full chain. In security, use separate identities and least-privilege permissions for each agent, enforce authorization in the systems and APIs they access, and log handoffs and tool calls. Those records help teams reconstruct activity, detect unintended privilege escalation, and revoke access when it is no longer needed.