When designing autonomous agents for enterprise environments, giving the AI the ability to string together actions (A to B to C) is incredibly powerful. However, true enterprise readiness requires guardrails. What happens when an agent designed to process 5,000 daily orders suddenly receives a request to process 1 million?

Blind autonomy in edge cases can consume massive resources, incur severe financial costs, and break downstream systems. This is where Human-in-the-Loop (HITL) architecture becomes critical.

In this post, we will dive deep into implementing multi-stage approval processes within your agentic workflows, ensuring your AI operates with the right balance of autonomy and human oversight.

Why Multi-Stage HITL?

HITL is essential when an agent requires clarification, additional context, or explicit authorization to proceed before an agent proceeds with an autonomous action. By injecting a multi-stage approval flow, you protect your infrastructure from outliers and errors (e.g., processing 1 million orders vs. 5,000) to save resources, effort, and money.

Common Enterprise Use Cases:

• Verifying procurement orders.
• Validating financial reports.
• Confirming project status.
• Escalating high-priority cases.
• Gathering human-only information.

---

Core Capabilities of Multi-Stage Approvals

Modern agent builder interfaces allow you to design complex routing logic without writing custom middleware. The core features support multiple configurations, including numerous approvers, stages, and conditions:

• Flexible Routing Logic: Choose between “First to respond” (workflow continues after one person acts) and “Everyone must approve” (workflow pauses until all approvers respond).
• Manual vs. AI Stages:
  • Manual Stage: Explicitly defined routing and assignees.
  • AI Stage: AI uses instructions to determine decisions (e.g., “reject if price > $1,000”, “approve if bill is attached”) and defines the next steps.
• Condition Sets: Adding statements/groups to trigger stages based on variables (equal to, greater than, less than).

---

Implementation Guide: The “Big Transaction Refund”

Let’s look at a practical scenario. A standard agent handles routine grocery refunds ($5–$15) autonomously. But if a user requests a refund for a high-value anomaly like a $2,000 refrigerator, we need a multi-approver HITL process.

Step 1: Establish the Trigger and Context Gathering

1. Navigate to your agent’s Topics and create a new one (e.g., “Big Transaction Refund”).
2. Define your trigger phrase (e.g., “big refund”).
3. Prompt the user by messaging them for transaction details: “Please provide transaction details, including amount and item.”
4. Capture this user input and transaction details into a variable (e.g., Var1) to pass into our approval flow.

Step 2: Build the Agent Flow

1. Add a tool and select New Agent Flow, then rename it appropriately.
2. In the designer, insert a new action and select Human in the Loop -> Run a multi-stage approval. (Note: This action is cancellable and pauses the workflow until completion).
3. Click Set up the approval process and create a Manual Stage (e.g., “Stage A”).

Step 3: Configure the Approval Stage

1. Select Type: Set the routing logic to “Everyone must approve”.
2. Assignees: Assign multiple users (e.g., Girish, Alex).
3. Context Injection: Dynamically pass context into the approval card. Include the customer name, URL, and item description.
4. Notification Settings: Enable email notifications and allow the reassignment feature so an approver can delegate the task.
5. Save the stage.

Step 4: Map the Outputs

Once the approval completes, the flow generates rich metadata. You need to map these workflow results back to your agent using variables to dictate the next steps.

1. Add output variables to your flow.
2. Select the Stage Outcome and Final Outcome.

3. Publish the flow.

Step 5: Finalize the Orchestration

1. Return to your main Topic.
2. Add the newly published flow as a tool.
3. Use the returned output variables to send a final status message to the user: “Based on the approval status, your request is: [Final Outcome]”.

---

Testing & Troubleshooting Tips

When testing and deploying HITL flows, keep these technical caveats in mind:

• The Connection Manager is Mandatory: Workflows will fail/complain if the connection to the HITL action isn’t established first. Ensure you open the Connection Manager and authenticate/establish the connection.
• Asynchronous Pausing: The agent state remains “processing” while waiting for “Everyone must approve” conditions to be met.
• Handling Timeout Errors: Flows can time out if approvers take longer than 2 minutes to respond during a live synchronous test.
• Flow Optimization: To handle timeouts and scale, use the flow’s edit settings to adjust execution parameters.
  • Look into Flow Capacity (allows a higher volume of actions) and Express Mode.

By effectively layering multi-stage approvals into your agents, you transition them from simple chatbots into robust, reliable enterprise automation engines.