Consider a hypothetical support agent processing a queue of refunds. One request points to an account that does not belong to the customer. Can you stop that request while keeping legitimate work available? Or does the session show a wider compromise that requires containment?
A control should support the response your situation needs. The label alone does not tell you whether it acts on a tool, an action, one agent or the whole workflow.
SCOPE
What does each control actually stop?
| Decision | Whole-agent containment | Per-action control |
|---|---|---|
| What is stopped? | The agent's continued operation or access, within the control's coverage. | A proposed action before its effect reaches the target. |
| When does it fit? | Continued execution cannot be trusted. | One action needs intervention and the remaining task can safely continue. |
| What happens to other work? | Queued and in-flight work may be interrupted. Recovery depends on checkpoints and downstream state. | Authorized work can continue after the task and session are assessed as trustworthy. |
| What evidence is required? | Proof of containment across the agent, outstanding work and relevant credentials. | A verdict linked to the action and proof of the actual target-system outcome. |
Some kill-switch implementations already provide capability-level or agent-level controls. Compare supported behavior and coverage rather than treating every alternative as an all-or-nothing shutdown. For definitions and implementation requirements, start with the AI agent kill-switch guide.
ONE ACTION AT A TIME
When should you block an action instead of the agent?
In the refund example, block the wrong destination before the payment runs. Preserve the case context and investigate why the request changed. A blocked action does not, by itself, establish that the next action is safe.
If the evidence supports a local mistake and the task remains trustworthy, a corrected proposal can receive a new decision. If the goal is hijacked, credentials are compromised or the agent is trying alternate paths around controls, contain the agent and investigate before resuming.
When I discuss the first case with our team, I call a whole-agent shutdown "overkill" for one bad move. The qualification matters: use the narrow response only when it contains the risk.
The refund is an illustrative scenario, not a reported customer incident or a completed product test.
DECISION CONTEXT
What makes Ackuity's control switch different?
Ackuity sits beside the agent in the execution path, outside the agent's control. It checks the action against the Agent Security Context Graph: 29 signals across the user, agent, intent and goal, target system and data, tools and supply chain, and history.
Deterministic policy checks enforce hard boundaries. Small language models interpret security context alongside rules and behavioral baselines. Within authorized scope, intent, history and risk determine execution or human review. No LLM judges the action.
A reviewer's approval cannot expand the agent's permissions. An authorized but unfamiliar bulk operation can require review; an operation that violates a hard policy limit remains blocked. Vinod's same-action context-graph example shows that distinction.
Ackuity checks 60+ threat models in 14 categories, mapped to NIST, OWASP and MITRE ATLAS. Decisions take 40 to 100 ms. That figure is decision time, separate from end-to-end overhead.
PROPORTIONATE RESPONSE
How does the response ladder work?
- Allow: signed, logged and run.
- Constrain: masked or scoped, then run.
- Human in the loop: paused for a person's decision.
- Block: dropped before the target.
- Terminate: container shut down, for catastrophic cases.
After a Block, do not silently retry the same action. After human review, changes to the target, amount or relevant context need a fresh decision. After Terminate, recovery requires a trusted execution state and an assessment of the affected credentials and downstream work.
COVERAGE
Which paths can the control enforce?
A control can enforce only the paths it covers. In Ackuity's routing model, 8 of 11 action types never reach a gateway. Your actual routing and credentials determine which boundary can stop a direct database call, shell command, queued task or agent hand-off.
The control belongs in the Goldilocks zone: not inside the agent, not at the gateway. Validate the boundary against the paths your agents actually use. The pre-execution implementation checklist explains what to check.
NVIDIA's Open Agent Safety Platform illustrates independent containment: OpenShell provides sandbox policy and NVIDIA describes Sentry as an out-of-band watchdog that can quarantine unsafe execution. Ackuity is an NVIDIA Inception member; our OpenShell supervisor integration is in development. Membership does not establish integration availability.
EVALUATE THE OUTCOME
What should a vendor demonstrate?
Use an isolated environment to observe an authorized action, an unauthorized action, a human-review boundary and whole-agent containment. Check the target-system record as well as the decision log. Then exercise alternate routes, queued work, child agents and a decision-service timeout.
Credential revoked, but work still running? Use the credential-revocation evaluation plan to separate a denied future request from an accepted job and a child agent’s own credentials.
Record supported paths, policy and product versions, legitimate-task completion, false positives, decision time and end-to-end overhead. The blank control-evaluation worksheet helps record those observations. It is a test plan, not evidence that a product has passed.
OWASP's Excessive Agency guidance recommends limiting privileges, requiring approval for consequential actions and enforcing authorization downstream. Use those principles alongside your containment and recovery tests.
Start with the 30-day Agent Exposure Assessment to examine existing telemetry and identify exposure. Inline enforcement and recovery behavior need a separate controlled evaluation.
NVIDIA, OpenShell and BlueField are trademarks of NVIDIA Corporation.
Putting human review into an execution path? Use the scoped-approval runbook and blank review record to define the operation, expiry, execution recheck and recovery evidence.