ML//agent//agent autonomy

Agent autonomy is the share of a trajectory that an agent may choose and execute without new human authorization, and it is set action by action, so an operator decides what the agent does alone and what waits for approval. An agent may autonomously browse and edit a sandbox while still requiring approval to spend money, contact a person or deploy.

Agent autonomy is the share of a trajectory that an agent may choose and execute without new human authorization, and it is set action by action, so an operator decides what the agent does alone and what waits for approval. An agent may autonomously browse and edit a sandbox while still requiring approval to spend money, contact a person or deploy.

Longer autonomy can compound useful work because the agent can investigate, recover and continue. It can compound mistakes for exactly the same reason. The important risk variable is therefore not autonomy alone, but autonomy multiplied by available tools, permissions, reachable environments and time. Agent containment works on the second factor: it caps the worst consequence of an autonomous trajectory without requiring the agent to behave perfectly.

Asking how long an agent can keep going, and what it can touch, is often more useful operationally than asking how intelligent it is.