HeadcountZero ← All entries

What is a blast radius?

The most an agent can damage with one action, or one bad run, before anything stops it.

Glossary3 min readUpdated 1 Aug 2026

Definition

blast radius (n.) – the largest amount of damage a single agent action, or a single bad run, can cause before something outside the agent stops it.

Blast radius is a ceiling, not a forecast. It does not ask what an agent is likely to do. It asks what the worst correct-looking mistake would reach: how much money could move, how many records could change, how many messages could go out under the company's name, and how many systems sit within reach of the credential the agent holds. The answer is set by the access the agent was given, not by the instructions it was given.

That distinction is why the term matters. An instruction that says to change at most ten prices limits nothing if the access allows the agent to change all of them. The blast radius is every price in the catalog, with a request attached. A cap enforced outside the model, a spend ceiling on the account, a batch limit in the tool, a credential scoped to one folder, is what actually draws the circle.

The term is borrowed and used loosely. In cloud security writing it usually means what an attacker could reach after a breach; Mitiga writes that modern blast radius is "defined by identity trust relationships and service permissions." In infrastructure engineering the same idea appears as fault isolation: AWS guidance on cell-based architecture describes splitting a workload into independent cells so that a failure such as a bad software update stays inside one cell. Here the term means the agent version: the reach of one action or one run, measured in the units the business cares about.

Blast radius is stated per procedure, in numbers. A weekly report has a blast radius of one wrong document a reader can ignore. A pricing job has a blast radius of every price it can write, multiplied by the hours until someone notices. Writing both down makes the difference visible, and that difference decides where a judgment gate belongs and which rung of the mandate ladder the procedure can hold.

It is reduced in three ways, and all three are code: narrower access, smaller batches, faster detection. Narrower access shrinks what one credential can touch. Smaller batches mean a bad run stops after the first batch rather than after all of them, because the verify step runs between batches. Faster detection shortens the time a mistake keeps compounding, which matters for anything on a schedule that repeats the same error on every pass. Detection only helps if something then ends the run, which is the job of a kill switch: halt the run, disable its schedule, revoke its access.

The common failure is measuring the action and forgetting the loop. One wrong update is small. The same wrong update applied by a job every hour for a weekend is not. Blast radius is the action multiplied by the repetitions before a person or a check intervenes.

Blast radius example

Say an agent drafts refunds for customer complaints. On propose, its blast radius is zero refunds issued, because a person applies each one; the cost of a wrong draft is the minute spent reading it. On apply within a cap of fifty dollars per refund and twenty refunds per day, the blast radius is one thousand dollars a day, plus whatever a wrong refund costs in trust. On apply with an account that can issue any refund, the blast radius is the balance of the account. The agent and its instructions are identical in all three cases. Only the access changed.

The useful question before raising a procedure's mandate is therefore not whether the agent is good enough. It is whether the business can absorb the full blast radius on the day the agent is wrong, because that day arrives for every procedure that runs long enough.

How caps fail

A cap fails when it lives in the wrong layer. A limit written in the prompt is read by the model and can be misread by it. A limit in the tool the agent calls holds for that tool and not for a second tool that reaches the same system. A limit set on the account itself, by the provider, holds for everything, which is why the outermost cap is the one to trust.

A cap also fails by going stale. Access granted for a small job tends to stay after the job grows, and the blast radius grows with it while the records still describe the old number. Reviewing access records on a schedule, which org as data makes cheap, is how the written number and the real one stay the same.