Shield Agent AI

All 13 questions

Question 8 of 13 · Permissions

Is human approval required before destructive or irreversible actions?

Clients also ask it like this:

  • Does a person approve before an agent deletes data or deploys?
  • What stops an agent from dropping a database?
  • Describe your human-in-the-loop controls for coding agents.

What the client is checking

The client wants to know if a second pair of eyes stands between an agent and an action that cannot be undone. They check if the step is a technical control or only a habit.

Three draft answers

Pick the level that is true for you today. Every statement marked [VERIFY] must be true for your organisation before you send it. Delete a statement that is not true. Fill each {{placeholder}}.

Not yet in place

Agents can take destructive actions without a second person approving them. [VERIFY] The developer watches the agent and can stop it. [VERIFY] We have not set a required approval step. [VERIFY] We will define one by {{date}}. [VERIFY]

Policy only

Our policy requires a person to approve destructive or irreversible actions by an agent: {{examples}}. [VERIFY] The developer who runs the agent is that person. [VERIFY] The step is a rule and a habit. No technical control forces it. [VERIFY]

Enforced, with evidence

A technical control stops these actions until a person approves them: {{classes}}. [VERIFY] The control stops shell commands and file actions, not only named tools. [VERIFY] The approver is {{role}} and an approval needs {{conditions}}. [VERIFY] Each approval or refusal is recorded with the time, the person and the action. [VERIFY] If the approval service is not available, the action does not run. [VERIFY] We tested this on {{date}} with {{test}}. [VERIFY]

Evidence to attach for the strongest answer

  • The definition of destructive and irreversible actions in your policy.
  • The configuration of the control.
  • A redacted record of one approval and one refusal.
  • The list of approvers.
  • The result of a test with the approval service down.

How to move up one level

From not yet in place to policy only

Define the destructive and irreversible actions for your stack (data deletion, force push, infrastructure destroy, production deploy). Write the rule that a named person approves each one before it runs.

From policy only to enforced, with evidence

Move the rule into a technical control that waits for the approval. Decide what happens when the approval service is down. The safe choice is that the action does not run. Test it, and keep the test record.

Where KYA fits

The local hook is advisory and fails open. It denies a fixed list of commands. A deny is a block, not an approval. Four commands are routine on a local database and risky on a remote one: DELETE FROM with no WHERE, prisma migrate reset, prisma db push --accept-data-loss and rm -rf of the project folder. On Claude Code the hook asks the person about these. Other hosts deny them.

For other calls the OSS client records a REQUIRE_APPROVE verdict as advisory and lets the call go on. --strict blocks such a call. It does not wait for a person. The hosted hook (preview) opens no approval request. A human approval path runs on the MCP route (kya.request_approval) and with kya wrap --hold. The call waits for a person, who approves or rejects with the kya.approve scope. This needs a hosted workspace. The hosted service fails closed. A busy approval service is never an approval: the call stops.

Limits of the hosted hook: the agent runs as the user. It can edit the settings or kill the hook. Removal is detected, not prevented. Only Claude Code has onFailure: block.

KYA does not know which of your actions are destructive beyond its rule lists and your workspace policy.

Get all 13 answers in one document

Choose the level that is true for each group. You get a preview at once and the full pack by email.

Make my Answer Pack

Previous question Next question

These are drafts. Send only what is true for your organisation. This is not legal advice.