HITL Overview
Allow your agent and users to collaborate on complex tasks.
What is this?
Human-in-the-loop (HITL) lets an agent pause mid-run to collect input, confirmation, or a choice from the user, then resume with that answer folded back into its reasoning. It's what turns an autonomous workflow into a collaborative one: the agent keeps its context, the user keeps the steering wheel.
When should I use this?
Use HITL when you need:
- Quality control: a human gate at high-stakes decision points.
- Missing information: the agent asks the user for a value it cannot find.
- Expert input: the user supplies domain knowledge the model lacks.
Two patterns for HITL in CopilotKit
useHumanInTheLoop: no frontend-tool code on the agent
AGUIStream turns every tool the client declares into a client-side tool,
so a tool you register with useHumanInTheLoop reaches the model through
the AG-UI request. Do not declare it on the AG2 agent. Name it in the
agent's prompt so the model calls it.
To keep the user's answer, have a backend tool write it to
context.variables. When a tool changes the variables, AGUIStream sends
a STATE_SNAPSHOT at the end of the run:
from typing import Annotated
from fastapi import FastAPI, Header
from fastapi.responses import StreamingResponse
from pydantic import Field
from ag2 import Agent, Context, tool
from ag2.ag_ui import AGUIStream, RunAgentInput
from ag2.config import OpenAIResponsesConfig
@tool
def store_user_choice(
choice: Annotated[str, Field(description="The option selected by the user")],
context: Context,
) -> str:
"""Store the latest user choice in shared state."""
context.variables["hitl"] = {"latest_choice": choice}
return f"Stored choice: {choice}"
agent = Agent(
name="assistant",
prompt=(
"When you need the user to choose, call the frontend tool `offerOptions`. "
"After it returns, call `store_user_choice` with the selected value."
),
config=OpenAIResponsesConfig(model="gpt-6-luna"),
tools=[store_user_choice],
)
stream = AGUIStream(agent)
app = FastAPI()
@app.post("/chat")
async def run_agent(
message: RunAgentInput,
accept: str | None = Header(None),
) -> StreamingResponse:
return StreamingResponse(
stream.dispatch(message, accept=accept),
media_type=accept or "text/event-stream",
)Approve a tool call with interrupts
AG2 can also pause a run on the backend. A tool wrapped in
ApprovalRequired is held until the user approves or rejects the call.
AG2 sends the held call to the client as an AG-UI interrupt, and
CopilotKit's useInterrupt renders it.
from typing import Annotated
from pydantic import Field
from ag2 import Context, tool
from ag2.middleware.builtin.tools.approval import ApprovalRequired
@tool(
description="Save a city to the user's favorites. The user confirms before anything is saved.",
middleware=[
ApprovalRequired("Save this city to your favorites?\n{tool_arguments}", allow_always=False)
],
)
async def save_favorite_city(
context: Context,
city: Annotated[str, Field(description="The city name to save, e.g. 'Lisbon'")],
) -> list[str]:
favorites = list(context.variables.get("favorites", []))
if city not in favorites:
favorites.append(city)
context.variables["favorites"] = favorites
return favoritesServe the agent with AGUIStream as usual. Do not set a hitl_hook on
the agent. With one, AG2 asks in-process and no interrupt reaches the
browser.
The run ends with an interrupt of reason tool_call. The backend reads the
answer as a boolean, so resolve(true) approves the call and
resolve(false) refuses it. A refusal does not fail the run: the model is
told the call was refused and carries on.
import { useInterrupt } from "@copilotkit/react-core/v2";
export function Page() {
useInterrupt({
render: ({ interrupt, resolve }) => (
<div>
<p>{interrupt?.message ?? "The assistant needs your confirmation."}</p>
<button onClick={() => resolve(true)}>Approve</button>
<button onClick={() => resolve(false)}>Reject</button>
</div>
),
});
// ...
}Ask Save Lisbon to my favorites. The approval card appears in the chat.
Approve it, and the city shows up in the shared state.
useInterrupt().resolve() and .cancel() send no metadata in the
resume entry. That is enough for a gated tool call like the one above,
but not for an AG2 server built with require_resume_proof=True, which
needs the interrupt's metadata copied into the resume entry. See
Asking the human a question
in the AG2 docs.
A held turn lives in the AG2 process that served the run, so the resuming run must reach that same process. See the AG2 deployment notes before you run several replicas.
CopilotKit ships two complementary ways to pause an agent turn and ask the human something. They look similar from the outside (the chat pauses, a custom component appears, the user answers, the run resumes), but they're wired differently on the backend, and each has its own niche.
| Pattern | Who decides to pause? | Backend surface |
|---|---|---|
useHumanInTheLoop | The LLM, by calling a registered client-side tool | A frontend-only tool description (Zod schema + render) |
useInterrupt | Your agent code, at a step that must stop for a person | An interrupt the agent emits, for example interrupt() in a LangGraph node |
Pick useHumanInTheLoop when the pause is an agent-initiated decision (the model chose to ask the user) and you want the picker UI inlined into the normal tool-call flow.
Pick useInterrupt when the pause is a checkpoint in your agent code: the code path always requires a human answer, and the server-side contract is an interrupt. On the Built-in Agent, interrupts are temporarily unavailable, so use useHumanInTheLoop there.
Pattern 1: useHumanInTheLoop (tool-based)
The app registers a HITL tool on the client with useHumanInTheLoop. When the LLM calls that tool, CopilotKit renders your Vue component and gives it a respond callback during the executing phase. The agent sees the answer as the tool result and continues from there.
<script setup lang="ts">
import { z } from "zod";
import { useHumanInTheLoop } from "@copilotkit/vue/v2";
import TimePickerCard from "./TimePickerCard.vue";
useHumanInTheLoop({
name: "schedule_meeting",
description: "Ask the user to select a meeting time.",
parameters: z.object({
topic: z.string(),
attendee: z.string(),
}),
render: TimePickerCard,
});
</script>
<template><p>Meeting approval tool registered.</p></template>TimePickerCard.vue receives the tool arguments and a respond callback. respond is defined only while the tool call is executing:
<script setup lang="ts">
const props = defineProps<{
args: { topic?: string; attendee?: string };
respond?: (result: unknown) => Promise<void>;
}>();
</script>
<template>
<section>
<p>Pick a time for {{ props.args.topic ?? "the meeting" }}</p>
<button
:disabled="!props.respond"
@click="props.respond?.({ chosen_time: '09:00' })"
>
Tomorrow at 09:00
</button>
</section>
</template>The picker UI can use static candidate slots, live availability, a calendar API, or any other application data. Mount the registration component beneath CopilotKitProvider so the composable can reach the runtime.
Pattern 2: useInterrupt (graph-paused)
With a runtime interrupt, your agent code enforces the pause:
- The agent interrupts with a serializable payload.
- The run stops.
- The client renders your UI with the payload.
- The user answers, and the client resumes the run with the answer.
CopilotKit's useInterrupt composable and CopilotChat's #interrupt slot form the Vue render contract.
See the useInterrupt deep dive for the full walkthrough, including the backend contract and Vue slot wiring.
Going headless
Both patterns above can drive custom UI. For runtime interrupts, set renderInChat: false and use the reactive state returned by useInterrupt; for a frontend-tool pause, let the Promise-returning HITL tool drive application state. See the headless interrupts guide for the interrupt path.
How is this guide?