
AI agent approval workflows: steps and examples
Build an AI approval workflow in six steps: choose actions that need review, assign approvers, set expiry, approve the exact request and confirm the result.
The decision in brief
An AI agent approval workflow pauses a proposed action for an authorized person to review. Choose which actions need review, assign an independent reviewer, set a deadline, bind the decision to the exact request and record what happened. In Durin, approval gives permission for the requester to retry that request before expiry; approval alone does not perform the action.
What an AI approval workflow is
An AI agent approval workflow determines when an agent must wait for a person before taking an action. For example, an agent can draft a customer update, while an assigned reviewer decides whether that specific message may be sent to that specific destination.
This is one form of human-in-the-loop review. It differs from using AI to automate an ordinary approval process, such as routing invoices or leave requests. It also differs from approving an AI tool for company use: an approved tool can still propose an action that needs separate review.
The MCP tools specification recommends user visibility and confirmation for sensitive operations. Durin adds independent review when its write policy requires it: the approvers an approval workflow names, or another active administrator, decide whether a specific request may proceed.
Six steps to build an AI agent approval workflow
Take an agent that prepares a customer-facing release update. Your team wants a release owner to review the destination and message before it is sent. In Durin, the policy determines whether that action needs approval; the approval workflow determines who can decide.
- Choose the trigger. Configure the policy so the relevant connection, tool and resource require review before an action can proceed.
- Assign reviewers and a deadline. Publish a workflow naming the people, teams or roles that can decide, with an escalation stage if needed. A deadline must not silently authorize the action.
- Submit the request with context. Confirm that the agent stops before sending anything. Identify the requester, destination, proposed message and consequences. Durin records an argument digest, so the reviewer also needs a trustworthy description of what those arguments would do.
- Record the decision. An assigned reviewer other than the requester chooses Allow once or Deny. Approval authorizes this exact request; it does not send the message.
- Retry the exact request. The requester retries before expiry. Changed arguments or authorization context require fresh evaluation; access is checked again.
- Confirm the outcome. Inspect the execution record and, if necessary, the destination system. Reconcile a timeout or uncertain result before attempting another action.
Try the sequence first with a supported sandbox action and two different people. Test denial, expiry, changed arguments and reviewer revocation as well as the successful path. Without a matching published workflow, a different active owner or administrator reviews the request.
Choose a workflow pattern
Approval workflows differ in who reviews, when they review and how much they see. Most teams combine more than one pattern, so that routine actions flow under rules and only consequential ones wait for a person.
Durin can route requests to named people, teams or roles in escalating stages, with review notifications in Slack when connected. Routed requests offer Allow once or Deny. Approval memory applies to eligible unrouted requests; it cannot bypass the people a published workflow assigns.
- Confirmation in the client: the person using the agent confirms each sensitive action. It is quick, but the requester reviews their own request.
- Independent review: a second person decides on the exact request. It is slower, but it separates requesting from approving.
- Review by exception: routine actions proceed under rules, and only unusual ones, such as large amounts or external recipients, go to a reviewer.
- Remembered decisions: an eligible repeated request can use a limited standing grant. This reduces review work but needs a deliberately narrow scope and revocation rules.
- Sampled review afterwards: someone checks a sample of completed actions. It helps with quality, but it cannot stop a harmful action.
Decide which actions need approval
Use the consequences of an action to guide review. An external message, bulk update, or payment can merit more scrutiny than a low-sensitivity lookup. Read-only tools can still disclose confidential information, so access checks remain necessary.
These questions are for your own risk assessment; Durin has no built-in risk tiers. In Durin, workspace defaults set the access and write decision, and custom rules can require approval for particular connections, tools, resources, environments, people or teams. Approval workflows then route each request to named people, teams or roles, in stages that escalate after a time limit. Without a matching workflow, any active administrator other than the requester reviews it.
- Impact: what data, money, or external communication could change?
- Scope: how many records or recipients could one call affect?
- Reversibility: how would you detect and repair a mistake?
- Review quality: does the reviewer have enough context to make a decision?
| Agent action | Review choice | Context to inspect |
|---|---|---|
| Look up a permitted sandbox ticket | Proceed under the configured access policy. | Requester, connection and permitted resource. |
| Send a customer-facing update | Require an independent reviewer before sending. | The destination, message, requester and exact request. |
| Change many customer records | Review the scope and recovery plan before authorizing. | Affected records, proposed changes and how to reconcile failures. |
Required checks come before approval
Durin first checks identity, membership, connection and team access, credentials, reviewed tools, resources, and inspection results. An approval cannot repair a failed mandatory check or expand upstream provider permissions.
Default allow applies only after required checks pass. Administrators can require independent review before agents make changes or deny those changes entirely; default deny blocks tool access. Sandbox and production use the same policy controls.
An unreviewed tool stays blocked pending tool review. Reviewing a tool definition and approving a particular write are separate decisions.
Bind the approval to the request
An exact grant covers the requester, organization, connection, credential context, tool and schema versions, resource, argument digest, policy version, and revocation state. Changing that context can invalidate the approval.
The argument digest is a fingerprint used to match the request; it is not a readable description of the operation. Durin’s review shows that digest alongside the requester, tool, resource, policy and schema versions, and expiry. Reviewers need the relevant business context before approving.
Grants expire and are single-use. Membership and authority are checked again before execution, so earlier approval does not preserve access after revocation.
Handle uncertainty without duplicating a write
A connection can fail after an upstream system has committed a change. No response does not mean no action. Approval, dispatch, and confirmed completion must remain distinct.
Durin records required execution intent before a governed write and does not automatically repeat a potentially completed write. Inspect the recorded outcome and reconcile it with the provider before deciding whether a new action is appropriate.
Single-use approval limits reuse of the grant; it is not a universal exactly-once guarantee across external systems. Include uncertain outcomes in your runbook and assign an owner to reconcile them.