Definition
A design principle that says an AI agent should be given the absolute minimum ability to take actions — and the fewest tools, permissions, and reach — needed to do its specific job, nothing more. It extends the familiar 'least privilege' idea from human user accounts to the broader concept of what an autonomous agent is actually allowed to do.
Why it matters
Agents with more autonomy than they need are the biggest source of damage when they are manipulated or malfunction; enforcing least agency limits the blast radius of any single compromised or confused AI agent.