The term comes from incident response, where it describes how far damage travels from a failure. Applied to agents it answers a specific question: if this thing does something nobody intended, what does it touch?
How to measure it
List the systems the agent can reach, not the ones it usually touches
List the actions it can take in each, written as verbs
Cost the worst chain those actions could form, not the worst single one
Note how fast each action could be undone, because reversibility sets the real cost
Where teams get it wrong
Measuring intended reach instead of granted reach. An agent's blast radius is defined by its permissions, not its job description, and those permissions were often inherited from a service account that accumulated access over years. Nobody would grant that whole pile in one sitting, which is why the honest measurement usually surprises the team that owns the agent.
A fast diagnostic: ask who is on call for a given agent, what its blast radius is, and whether they can stop it inside a minute. Blank looks on any of those point at an operational gap rather than a technical one.