Guides  /  Governance and control

How do you stop an agent before it acts?

Answer

Stopping an action requires running in the same path that executes it. A check that reads the result afterwards can only record what already happened.

Akka Verify runs guardrails inline in the transactional path. The check finishes before a response is emitted or a tool call is made, and the action, the verdict and the audit entry are written in one transaction.

The mechanism

Position in the request path decides what a check can do.

A check runs either outside the request path or inside it. Outside, the check receives a copy of what the agent did. The check can score the response, raise an alert, and open a ticket. The action it is scoring has already executed.

Inside, the check holds the request while it runs. The agent proposes an action, the guardrail returns a decision, and the runtime acts on that decision before anything reaches a customer or a system of record.

INSIDE THE REQUEST PATHagentproposesguardraildecides before executionallowedthe action runsblockedheld for a personOUTSIDE THE REQUEST PATHobserver
What an inline check can do

The verdict allows, blocks, sanitizes, or escalates.

AllowThe action executes. The verdict and the action are recorded together, so the log shows the check ran and what it concluded.
BlockThe action does not execute. The agent receives the refusal and can be re-prompted with the constraint that failed.
SanitizeA sanitizer redacts, masks, or reshapes the content before it continues. The original and the change are both recorded.
EscalateThe run pauses and a person decides. Durable suspension holds the paused run across crashes and deployments for as long as the decision takes.
The cost

An inline check adds latency to every request it guards.

A check in the request path runs before the response, so its duration is added to the response. The latency cost is why each control chooses its own mode.

Compliance owns the policy matrix that makes the choice. A control that must prevent an outcome runs inline. A control that only has to be evidenced runs against the interaction log after the fact. Inline checks and after-the-fact checks produce the same evidence record. Only the inline check can change what happens.

Questions

Related questions.

What happens if the guardrail itself fails?

The runtime treats an unavailable check as a failed check, so the action does not execute. A guardrail that fails open is a guardrail that stops enforcing under exactly the load that makes enforcement matter.

Does a human review stall the system?

The paused run holds its state and releases its resources. Other work continues, and the held run resumes at the step where it stopped once the decision arrives, whether that takes a minute or three days.

Who decides which controls block?

Compliance owns the policy matrix, and the runtime enforces it. Engineers do not hand-code the decision, so a change to policy does not require a change to application code.

What does the record show when an action is blocked?

The proposed action, the policy that refused it, the verdict, and the time. The record is written in the same transaction as the refusal, so an auditor reads the decision itself. A missing result would leave the auditor inferring one.

Can this run against agents built elsewhere?

Evaluation and evidence work against agents built on other frameworks. Inline enforcement requires the runtime to sit in the execution path, which means the agent runs on Akka.

On Akka. Guardrails, sanitizers, durable human-in-the-loop suspension and the interaction record are capabilities of Akka Verify.