Human Approval for AI Agents: Bind It to the Action
Design human approval for AI agents around concrete proposals, verified inputs, expiry, and execution checks. Avoid vague confirmations and duplicate side effects.
Human approval for AI agents is meaningful only when the person can see what they are authorizing and the executor performs that same action. A button labeled “Continue” is weak protection if the destination, amount, or document contents can change afterward.
Consider an assistant preparing an external email. The user approves a draft for one recipient. Before execution, a later model step adds another recipient and an attachment. Reusing the earlier approval would authorize something the user never reviewed.
Decide which actions need review
Classify actions by their effects, reversibility, and scope. Reading an already authorized document, drafting a local response, publishing a public statement, and changing access permissions should not inherit the same approval policy.
Use deterministic policy where possible. The application can know whether an operation sends data externally, deletes records, or changes a protected setting. Do not rely solely on a generated risk score from the model proposing the action.
A practical policy may allow routine reads automatically, require review for external messages, and prohibit certain operations entirely. Those are examples, not universal defaults. The correct boundaries depend on the product and the authority the user has actually delegated.
Avoid asking for approval so often that users stop inspecting it. Make low-impact work useful and automatic where authorized, while reserving meaningful review for decisions with consequences.
Create an immutable proposal
Store the proposed action server-side before showing the approval interface. Include the operation type, target resources, significant parameters, relevant content, creator, and creation time. Give it a stable identifier and version.
For the email example, significant fields include sender identity, recipients, subject, body, and attachments. For a permission change, they include the resource, principal, and exact access being granted. A summary alone may omit the field that matters most.
OWASP's transaction authorization guidance emphasizes that users must identify and acknowledge significant transaction data. Applied to an agent workflow, that means the review interface should display the actual action being authorized.
If the proposal changes, invalidate the old approval or create a new version requiring review. A hash can help identify a specific payload, but it does not replace authenticated approval storage or clear presentation to the user.
Make the approval record explicit
Record who approved which proposal version, when, and under what policy. Check that the approver has authority over the affected resource. The fact that someone can view a draft does not necessarily mean they can publish it or grant access.
Define expiration. A proposal to modify an order may become stale when the order ships. A deployment approval may depend on a particular build artifact. Time alone is not the only validity condition; changes to relevant state may also require another review.
An illustrative state machine is:
proposed -> approved -> executing -> completed
| | |
rejected expired failed
Add states only when they have clear operational meaning. “Approved” should not imply “executed,” and “failed” should not imply that no side effect happened. The execution result may need reconciliation with the external system.
Recheck immediately before execution
The executor should load the approved proposal from trusted storage, validate its version and expiry, recheck relevant permissions, and confirm that the target state still satisfies the operation's preconditions.
Do not accept a fresh model-generated payload together with an approval ID and assume they match. Execute the stored approved payload or perform an exact comparison under a well-defined serialization scheme.
If the order status changed, stop with a stale-state result. If the recipient list changed, require approval of the new list. These checks belong at the action boundary, even if earlier stages already validated the proposal.
Keep the executor's credentials scoped to the operation. Human review is not a reason to give the agent unrestricted access to unrelated systems.
Design for retries and interruptions
A network timeout can leave the application uncertain whether an external action completed. Use a stable operation identifier and the external service's idempotency support where available. If the service lacks it, build a reconciliation strategy before allowing automatic retries.
LangGraph's interrupt documentation notes that resumed nodes can rerun work before an interrupt and recommends idempotent side effects or separating them from the pause. That is a concrete reminder to understand your workflow engine's replay behavior.
Place the approval pause before the irreversible operation. Do not send the email and then ask whether the user approves sending it. Also test process restarts between approval, execution, and result recording; those are common places for duplicate actions or lost status.
Show outcomes the user can verify
After execution, report what actually happened using the external system's result where possible. Include an appropriate reference, such as a published URL or operation ID. If the result is uncertain, say that it is being reconciled rather than claiming success.
For a rejection, preserve the user's reason when useful and avoid repeatedly presenting the same unchanged proposal. For a revision, show the changed fields so review remains focused.
Test stale approvals, modified payloads, unauthorized approvers, duplicate submissions, revoked access, and ambiguous external failures. These tests should exercise application code directly, then the agent experience around it.
The tool-permission guide defines what the approved executor is allowed to do. The idempotency design explains how retries can avoid repeating a side effect after approval. Bind consent, execution, and the recorded outcome to the same concrete proposal.
A well-designed approval boundary gives the user a concrete decision and gives the executor a precise contract. The agent can prepare the work, but the approved action remains stable from review through execution.