Add Human-in-the-Loop Controls to an Agent SDK Agent

Goal: Add human-in-the-loop (HITL) controls to an existing Agent SDK agent so one of its tools can auto-resolve routine decisions and pause for human input on high-stakes ones.

Outcome: Your existing callModel loop keeps running normally for routine tool calls, pauses with status: 'awaiting_hitl' for high-stakes calls, surfaces the pending call to your UI or API, and resumes after a human supplies the tool result.

You can give this page to your coding agent as the implementation brief. It should adapt the example names, storage, threshold, and user-review surface to your existing agent rather than scaffold a separate app.

HITL vs requireApproval

Both pause for human input, but they solve different problems:

Use HITL when the decision depends on the input data. Use requireApproval when you need a human to approve whether a tool should execute. See the Tool Approval & State reference for details on approval flows and conditional approval predicates.

Prerequisites

• An existing TypeScript agent that uses @openrouter/agent and callModel
• An OpenRouter API key configured in that agent’s environment
• A StateAccessor or a place to persist conversation state
• A UI, CLI, queue, or API surface where a human can review pending calls

1. Choose the tool that needs HITL

Pick the tool in your existing agent where the result sometimes needs human judgment. In this example, the agent can approve small payments automatically but must pause before approving larger ones.

A HITL tool uses onToolCalled instead of execute. The hook receives the parsed input and decides per-call whether to return a tool result immediately or pause for a human.

Return a value to auto-resolve (like a regular tool). Return null to pause the loop — the conversation status moves to 'awaiting_hitl' and the call surfaces to the caller.

1
import { OpenRouter, tool } from '@openrouter/agent';
2
import type { ConversationState, StateAccessor } from '@openrouter/agent';
3
import { z } from 'zod';
4
5
const paymentInputSchema = z.object({
6
  amount: z.number(),
7
  recipient: z.string(),
8
});
9
10
const paymentDecisionSchema = z.object({
11
  approved: z.boolean(),
12
  reviewedAt: z.number().optional(),
13
});
14
15
const approvePayment = tool({
16
  name: 'approve_payment',
17
  description: 'Approve a payment, escalating large amounts to a human',
18
  inputSchema: paymentInputSchema,
19
  outputSchema: paymentDecisionSchema,
20
  onToolCalled: async (input) => {
21
    if (input.amount < 100) {
22
      return { approved: true };
23
    }
24
    // Pause for human review
25
    return null;
26
  },
27
});

outputSchema is required for HITL tools — it validates both the auto-resolved return value and any human-supplied response. See the HITLTool type reference for the full type signature.

2. Add post-processing with onResponseReceived

When a human supplies a response for a paused call, onResponseReceived fires before the result reaches the model. Use it to enrich, validate, or transform the raw human input.

Use onResponseReceived when the human review surface does not return the exact model-facing tool result you want. Common cases include:

• Adding audit metadata such as reviewedAt, reviewerId, or an internal approval ticket ID
• Normalizing UI form fields into the tool’s outputSchema
• Validating the human response against your own policy before the model sees it
• Converting an approval, rejection, or edited value into the final tool result

Replace the tool definition from step 1 with this version:

1
const approvePayment = tool({
2
  name: 'approve_payment',
3
  description: 'Approve a payment, escalating large amounts to a human',
4
  inputSchema: paymentInputSchema,
5
  outputSchema: paymentDecisionSchema,
6
  onToolCalled: async (input) => {
7
    if (input.amount < 100) {
8
      return { approved: true };
9
    }
10
    return null;
11
  },
12
  onResponseReceived: async (raw) => {
13
    // Normalize and validate the human review result before it is sent back to
14
    // the model as the tool output. This is where you would add reviewer
15
    // metadata, enforce policy, or adapt UI fields to the tool's output schema.
16
    const decision = paymentDecisionSchema.parse(raw);
17
18
    return { ...decision, reviewedAt: Date.now() };
19
  },
20
});

If parsing or onResponseReceived throws, the error is surfaced to the model as { error: ..., originalOutput: ... }. If omitted, the human-supplied value passes through directly after schema validation.

3. Add it to your callModel loop and detect a pause

Add the HITL tool to the tools array you already pass to callModel. HITL resume requires conversation state, so reuse your existing StateAccessor or add one if your agent is currently stateless.

The snippet below shows the minimum shape with in-memory state for clarity. In production, back the StateAccessor with your database, Redis, or whatever storage your agent already uses.

1
// Keep using your existing OpenRouter client if your agent already has one.
2
const openrouter = new OpenRouter({
3
  apiKey: process.env.OPENROUTER_API_KEY,
4
});
5
6
// Use your existing tools array and StateAccessor if you already have them.
7
const tools = [approvePayment] satisfies readonly [typeof approvePayment];
8
const store = new Map<string, ConversationState<typeof tools>>();
9
const conversationId = 'conv-1';
10
const state: StateAccessor<typeof tools> = {
11
  load: async () => store.get(conversationId) ?? null,
12
  save: async (s) => { store.set(conversationId, s); },
13
};
14
15
const result = openrouter.callModel({
16
  model: '~openai/gpt-latest',
17
  input: 'Pay $500 to Acme Corp for the May invoice',
18
  tools,
19
  state,
20
});

If you need a deterministic smoke test, temporarily force this tool call:

1
toolChoice: { type: 'function', name: 'approve_payment' },

In production, your agent instructions or user request can let the model decide when to call the tool.

When the model invokes the tool and onToolCalled returns null, the result pauses with status: 'awaiting_hitl'. Check the state after the call completes, then surface the pending call to the human review surface in your app.

1
const stateSnapshot = await result.getState();
2
const pendingCalls =
3
  stateSnapshot.status === 'awaiting_hitl'
4
    ? await result.getPendingToolCalls()
5
    : [];
6
7
for (const call of pendingCalls) {
8
  console.log(`Pending: ${call.name}(${JSON.stringify(call.arguments)})`);
9
  console.log(`Call ID: ${call.id}`);
10
}

Illustrative output shape:

Pending: approve_payment({"amount":500,"recipient":"Acme Corp"})
Call ID: call_abc123

4. Resume with human input

Collect the human’s decision and resume by calling callModel again with a function_call_output item for each paused call.

In the payment example, the human review surface could be as simple as:

• An admin page with Approve and Reject buttons for the pending payment
• A Slack or Discord message where an operator clicks an approval action
• A CLI prompt that asks an internal user to confirm the payment
• A queue worker that waits for a back-office system to write the approval result

In other HITL workflows, the human input might be more than a boolean. A support escalation tool might collect an edited reply, a deployment tool might collect a rollback plan, or a data-change tool might collect corrected field values. Whatever collects the input should return a value that matches the tool’s outputSchema.

1
// Simulate collecting a human decision
2
const humanDecision = { approved: true };
3
const firstPendingCall = pendingCalls[0];
4
5
if (!firstPendingCall) {
6
  throw new Error('No pending HITL call to resume');
7
}
8
9
const resumed = openrouter.callModel({
10
  model: '~openai/gpt-latest',
11
  input: [
12
    {
13
      type: 'function_call_output',
14
      callId: firstPendingCall.id,
15
      output: JSON.stringify(humanDecision),
16
    },
17
  ],
18
  tools,
19
  state,
20
});
21
22
const text = await resumed.getText();
23
console.log(text);

The onResponseReceived hook fires on the human-supplied output before the model sees it. In this example, it adds a reviewedAt timestamp.

Check your work

• Calls below the threshold auto-resolve without pausing the loop.
• Calls above the threshold pause with status: 'awaiting_hitl'.
• pendingToolCalls contains the paused call with its id and arguments.
• Resuming with a function_call_output item continues the conversation.
• onResponseReceived transforms the human response before the model sees it.
• Changing the threshold or adding new conditions in onToolCalled does not require changes to the resume flow.