Pausing the Agent for Input

Pause an agent run mid-tool, hand control to a custom React component, and resume with the user's answer.

Human in the Loop: Interrupts

Waking up the demo

What is this?

Some agent steps need a person's answer before the run can continue, such as approving a payment or picking a meeting time. useInterrupt pauses the run, shows your React component, and resumes the run with the answer. How the pause works depends on your agent framework.

This framework ships a first-class interrupt primitive that lets running work suspend itself and hand control to the client: interrupt() in LangGraph, interrupts in AWS Strands, a tool that returns suspend() in Mastra, or a @human_feedback step in a CrewAI Flow. The run is frozen server-side until the client resolves the interrupt with a payload. Then the run resumes, and the agent receives that payload as the answer.

CopilotKit's useInterrupt is the frontend half of that contract: it subscribes to the paused run, renders whatever component you give it, and calls the agent back with the user's answer.

When should I use this?

Reach for useInterrupt when the pause is a checkpoint in your agent code where the code path must stop and wait for a human, not an LLM-initiated tool call. Typical cases:

  • A sensitive action (payments, irreversible writes) must be approved
  • A required piece of state isn't known and can only be collected from the user
  • The agent explicitly reaches an approval node in a longer workflow
  • You want the server-side contract to be an interrupt that resumes with a payload

For LLM-initiated pauses where the model decides on the fly to ask the user, prefer useHumanInTheLoop.

The backend: pause the run in your agent

Install the CrewAI AG-UI package

pip install ag-ui-crewai litellm

Give useHumanInTheLoop tools to the model

A Flow makes its own model call, so it must pass the tools itself:

  • Type the Flow on CopilotKitState. A tool registered with useHumanInTheLoop arrives in self.state.copilotkit.actions.
  • Pass the actions as tools, inside copilotkit_stream, so the tool call reaches the browser as it streams.
  • The run ends when the model calls the tool. After the user answers, CopilotKit runs the Flow again with the answer as the last message, with role: "tool". Read it in the @start() step to decide what happens next.
agent.py
from typing import Any, cast

from crewai.flow.flow import Flow, listen, start
from litellm import acompletion

from ag_ui_crewai import CopilotKitState, copilotkit_stream


class AgentState(CopilotKitState):
    pass


class SampleAgentFlow(Flow[AgentState]):
    @start()
    async def check_for_user_feedback(self):
        if not self.state.messages:
            return

        last_message = cast(Any, self.state.messages[-1])

        # The result of the useHumanInTheLoop tool (SEND or CANCEL)
        if last_message["role"] == "tool":
            if last_message.get("content") == "SEND":
                self.state.messages.append({
                    "role": "assistant",
                    "content": "✅ Great! Sending your essay via email.",
                })
                return

            if last_message.get("content") == "CANCEL":
                self.state.messages.append({
                    "role": "assistant",
                    "content": "❌ Okay, we can improve the draft. What would you like to change?",
                })
                return

    @listen(check_for_user_feedback)
    async def chat(self):
        system_message = {
            "role": "system",
            "content": (
                "You write essays. Use your tools to write an essay; "
                "don't just write it in plain text."
            ),
        }

        response = await copilotkit_stream(
            await acompletion(
                reasoning_effort="none",
                allowed_openai_params=["reasoning_effort"],
                model="openai/gpt-6-luna",
                messages=[system_message, *self.state.messages],
                tools=self.state.copilotkit.actions,
                stream=True,
            )
        )

        self.state.messages.append(response.choices[0].message)

Pause the Flow for useInterrupt

To pause at a fixed step in your code instead, decorate the step with CrewAI's @human_feedback and the AG-UI feedback provider. The step's return value is the interrupt payload. The listening step receives the user's answer as a HumanFeedbackResult.

src/agents/interrupt_flow.py
import json
from typing import Any

from crewai.flow import Flow, HumanFeedbackResult, human_feedback, listen, start

from ag_ui_crewai import CopilotKitState, agui_feedback_provider


class InterruptFlow(Flow[CopilotKitState]):
    @start()
    @human_feedback(
        message="Choose a meeting time or cancel the request.",
        llm=None,
        provider=agui_feedback_provider,
    )
    async def request_schedule(self) -> dict[str, Any]:
        # ... call the model, then return what the picker needs:
        return {"topic": topic, "attendee": attendee, "slots": slots}

    @listen("request_schedule")
    async def confirm_schedule(self, result: HumanFeedbackResult) -> None:
        feedback_text = result.feedback or ""
        # A cancelled interrupt resumes with an empty feedback string.
        cancelled = not feedback_text.strip()
        feedback = {} if cancelled else json.loads(feedback_text)

Register the endpoint so it emits standard AG-UI interrupts:

src/agent_server.py
add_crewai_flow_fastapi_endpoint(
    app,
    interrupt_flow,
    "/interrupt",
    emit_interrupt_outcome=True,
    enable_legacy_on_interrupt_event=False,
)

For a Conversational Flow, pass these two flags in addition to conversational=True. On the frontend, read the payload from interrupt.metadata.crewai.output. See src/agents/interrupt_flow.py for the full Flow.

In the tool or step that needs the user's answer, call your framework's interrupt primitive with the data your UI needs: interrupt() in LangGraph, suspend() in Mastra, or a @human_feedback step in a CrewAI Flow. The example agent does this in its schedule_meeting tool:

interrupt_flow.py
class InterruptState(CopilotKitState):
    pending_tool_call_id: str | None = None
    meeting: dict[str, Any] = Field(default_factory=dict)


class InterruptFlow(Flow[InterruptState]):
    """Pause after schedule_meeting and continue from a spec resume entry."""

    @start()
    @human_feedback(
        message="Choose a meeting time or cancel the request.",
        llm=None,
        provider=agui_feedback_provider,
    )
    async def request_schedule(self) -> dict[str, Any]:
        response = await copilotkit_stream(
            await acompletion(
                allowed_openai_params=["reasoning_effort"],
                reasoning_effort="none",
                model="openai/gpt-6-luna",
                messages=[
                    {"role": "system", "content": SYSTEM_PROMPT},
                    *self.state.messages,
                ],
                tools=[SCHEDULE_MEETING_TOOL],
                tool_choice={
                    "type": "function",
                    "function": {"name": "schedule_meeting"},
                },
                parallel_tool_calls=False,
                stream=True,
            )
        )
        message = response.choices[0].message
        self.state.messages.append(message)
        tool_calls = message.get("tool_calls") or []
        schedule_call = next(
            (
                call
                for call in tool_calls
                if call.get("function", {}).get("name") == "schedule_meeting"
            ),
            None,
        )
        if schedule_call is None:
            return {
                "topic": "Meeting",
                "attendee": None,
                "slots": candidate_slots(),
            }

        try:
            arguments = json.loads(
                schedule_call.get("function", {}).get("arguments") or "{}"
            )
        except (TypeError, json.JSONDecodeError):
            arguments = {}
        self.state.pending_tool_call_id = schedule_call.get("id")
        slots = arguments.get("slots")
        return {
            "topic": arguments.get("topic") or "Meeting",
            "attendee": arguments.get("attendee"),
            "slots": slots if isinstance(slots, list) and slots else candidate_slots(),
        }

    @listen("request_schedule")
    async def confirm_schedule(self, result: HumanFeedbackResult) -> None:
        feedback_text = result.feedback or ""
        try:
            feedback = json.loads(feedback_text or "{}")
        except (TypeError, json.JSONDecodeError):
            feedback = {}
        if not isinstance(feedback, dict):
            feedback = {}
        output = result.output if isinstance(result.output, dict) else {}
        # The AG-UI CrewAI provider resumes a cancelled protocol interrupt
        # with an empty feedback string. Submitted choices are JSON objects,
        # so an empty payload is the package's authoritative cancel signal.
        cancelled = not feedback_text.strip() or bool(feedback.get("cancelled"))
        self.state.meeting = {
            "topic": output.get("topic") or "Meeting",
            "attendee": output.get("attendee"),
            "time": None if cancelled else feedback.get("chosen_time"),
            "label": None if cancelled else feedback.get("chosen_label"),
            "cancelled": cancelled,
        }

        if self.state.pending_tool_call_id:
            result_content = json.dumps(self.state.meeting)
            self.state.messages.append(
                {
                    "role": "tool",
                    "tool_call_id": self.state.pending_tool_call_id,
                    "content": result_content,
                }
            )
            await copilotkit_emit_tool_result(
                self.state.pending_tool_call_id, result_content
            )
        response = await copilotkit_stream(
            await acompletion(
                allowed_openai_params=["reasoning_effort"],
                reasoning_effort="none",
                model="openai/gpt-6-luna",
                messages=[
                    {
                        "role": "system",
                        "content": (
                            f"{SYSTEM_PROMPT}\nScheduling outcome: "
                            f"{json.dumps(self.state.meeting)}"
                        ),
                    },
                    *self.state.messages,
                ],
                tools=[SCHEDULE_MEETING_TOOL],
                parallel_tool_calls=False,
                stream=True,
            )
        )
        self.state.messages.append(response.choices[0].message)

The run stops here until the client resolves the interrupt. The resolve payload comes back to the agent as the answer: the return value of interrupt() in LangGraph, resumeData in Mastra, or the HumanFeedbackResult that the next step of a CrewAI Flow receives. The example turns that value into a string for the model.

Two things to note:

  • The payload ({"topic": topic, "attendee": attendee}) is what the frontend reads off the interrupt. Keep it a plain, serializable object. It's the "pause-time context" the UI needs to render.
  • The return-side contract ({chosen_label, chosen_time} or {cancelled: true}) is entirely yours. The client can send anything as the resolve payload; the tool is the one that gives it meaning.

The frontend: useInterrupt render prop

On the client you register a useInterrupt hook per agent. When the paused run arrives, render receives both the raw event and the interrupt itself, and resolve(...) is how you resume the run. Your agent framework decides where the payload sits:

  • Custom-event interrupt (on_interrupt): event.value.
  • Standard AG-UI interrupt: interrupt.metadata, or a JSON string in interrupt.message.
page.tsx
import {
  CopilotKit,
  CopilotChat,
  useInterrupt,
} from "@copilotkit/react-core/v2";
import type { TimeSlot } from "./_components/time-picker-card";
import { TimePickerCard } from "./_components/time-picker-card";
import { generateFallbackSlots } from "../_shared/interrupt-fallback-slots";
import { useGenUiInterruptSuggestions } from "./suggestions";

type SchedulingPayload = {
  topic?: string;
  attendee?: string;
  slots?: TimeSlot[];
};

export default function GenUiInterruptDemo() {
  return (
    <CopilotKit runtimeUrl="/api/copilotkit" agent="gen-ui-interrupt">
      <div className="flex justify-center items-center h-screen w-full">
        <div className="h-full w-full max-w-4xl">
          <Chat />
        </div>
      </div>
    </CopilotKit>
  );
}

function Chat() {
  useGenUiInterruptSuggestions();
  useInterrupt({
    agentId: "gen-ui-interrupt",
    renderInChat: true,
    render: ({ event, interrupt, resolve }) => {
      const metadata = interrupt?.metadata as
        | { crewai?: { output?: SchedulingPayload } }
        | undefined;
      const raw = event.value ?? {};
      const fallback = (typeof raw === "string" ? JSON.parse(raw) : raw) as
        | SchedulingPayload
        | { metadata?: { crewai?: { output?: SchedulingPayload } } };
      const payload: SchedulingPayload =
        metadata?.crewai?.output ??
        ("metadata" in fallback
          ? fallback.metadata?.crewai?.output
          : (fallback as SchedulingPayload)) ??
        {};
      const slots =
        payload.slots && payload.slots.length > 0
          ? payload.slots
          : generateFallbackSlots();
      return (
        <TimePickerCard
          topic={payload.topic ?? "a call"}
          attendee={payload.attendee}
          slots={slots}
          onSubmit={(result) => setTimeout(() => resolve(result), 500)}
        />
      );
    },
  });

  return (
    <CopilotChat agentId="gen-ui-interrupt" className="h-full rounded-2xl" />
  );
}

Whatever you pass to resolve is sent back to the agent as the answer to the matching interrupt, for example the return value of interrupt(...) in LangGraph.

Ask the agent to book a call. The time picker appears in the chat. Pick a slot, and the agent confirms the booked time.

Key props

  • agentId: must match a runtime-registered agent. If omitted, the hook uses the agent of the surrounding chat configuration, then the agentId set on CopilotKitProvider, then "default". A mismatch means the interrupt never fires.
  • render: receives { event, interrupt, resolve }.
  • renderInChat: when true (as above), the picker appears inline in the chat transcript, between the paused assistant turn and the still-pending continuation. To show it outside the chat, see Headless interrupts.

Multiple interrupts? Add a type and gate with enabled

If your agent issues more than one kind of interrupt (e.g. "ask" vs "approval"), tag each with a type field on the payload and install one useInterrupt per shape, each gated by an enabled predicate:

// The LangGraph adapter sends `event.value` as a JSON string.
const readValue = (value: unknown) =>
  typeof value === "string" ? JSON.parse(value) : value;

useInterrupt({
  agentId: "gen-ui-interrupt",
  enabled: (event) => readValue(event.value).type === "ask",
  render: ({ event, resolve }) => (
    <AskCard question={readValue(event.value).content} onAnswer={resolve} />
  ),
});

useInterrupt({
  agentId: "gen-ui-interrupt",
  enabled: (event) => readValue(event.value).type === "approval",
  render: ({ event, resolve }) => (
    <ApproveCard content={readValue(event.value).content} onAnswer={resolve} />
  ),
});

Preprocess with handler

For cases where the interrupt can sometimes be resolved without user input (e.g. the current user already has permission), pass a handler that runs before render. The handler can call resolve(...) itself to resume the agent early. The interrupt card unmounts when the resume run starts. Or return a value that render receives as result:

useInterrupt({
  agentId: "gen-ui-interrupt",
  handler: async ({ event, resolve }) => {
    const value =
      typeof event.value === "string" ? JSON.parse(event.value) : event.value;
    const dept = await lookupUserDepartment();
    if (value.accessDepartment === dept || dept === "admin") {
      resolve({ code: "AUTH_BY_DEPARTMENT" });
      return; // agent will resume; card unmounts when the run starts
    }
    return { dept };
  },
  render: ({ result, event, resolve }) => (
    <RequestAccessCard
      dept={result?.dept}
      onRequest={() => resolve({ code: "REQUEST_AUTH" })}
      onCancel={() => resolve({ code: "CANCEL" })}
    />
  ),
});

AG-UI standard interrupt flow vs. custom-event flow

This section is a reference for developers who write render code for their own agent. The useInterrupt reference lists every option. useInterrupt handles two kinds of interrupt, and your agent framework decides which kind it sends.

Standard flow (RUN_FINISHED with outcome.type === "interrupt")

When the agent backend conforms to the AG-UI protocol, it signals an interrupt by emitting a RUN_FINISHED event whose outcome carries the interrupts array:

outcome.type === "interrupt"
outcome.interrupts  // Interrupt[]

The hook detects this on onRunFinishedEvent and exposes the interrupts on the render props after onRunFinalized fires. Your render function receives:

  • interrupt: the primary Interrupt (interrupts[0]), with shape { id, reason, message?, toolCallId?, responseSchema?, expiresAt?, metadata? }.
  • interrupts: the full open set (usually one, but multi-interrupt is supported, as described below).
  • resolve(payload?, interruptId?): records { status: "resolved", payload } for the targeted interrupt (defaults to the primary). The agent run resumes once every open interrupt has a response.
  • cancel(interruptId?): records { status: "cancelled" } for the targeted interrupt. Same accumulate-then-submit logic applies.

Each resume entry carries the interrupt id, the status, and your payload. resolve and cancel send no metadata.

Custom-event flow (on_interrupt)

Agents that emit a custom on_interrupt event use this flow. The LangGraph adapter emits this event by default. The hook detects this on onCustomEvent and sets interrupt to null and interrupts to []. The payload is in event.value. Calling resolve(payload) resumes via forwardedProps.command (the custom-event resume mechanism). cancel() dismisses the interrupt without resuming. The agent never receives a response.

Priority

If both signals appear on the same run, the standard flow wins.

Approve / Cancel example

function ApprovalInterrupt() {
  useInterrupt({
    render: ({ interrupt, resolve, cancel }) => (
      <div className="p-3 border rounded">
        <p>{interrupt?.message ?? "Approve this action?"}</p>
        <div className="mt-2 flex gap-2">
          <button onClick={() => resolve({ approved: true })}>Approve</button>
          <button onClick={() => cancel()}>Cancel</button>
        </div>
      </div>
    ),
  });
  return null;
}

resolve({ approved: true }) records a resolved entry and submits the resume array to the agent. cancel() records a cancelled entry and does the same. Both return the RunAgentResult once the run restarts.

Multi-interrupt behavior

Some agents issue more than one interrupt in a single run (e.g. two independent approvals). Each interrupt has its own id. Address them individually:

useInterrupt({
  render: ({ interrupts, resolve, cancel }) => (
    <ul>
      {interrupts.map((i) => (
        <li key={i.id}>
          {i.message}
          <button onClick={() => resolve({ ok: true }, i.id)}>Approve</button>
          <button onClick={() => cancel(i.id)}>Cancel</button>
        </li>
      ))}
    </ul>
  ),
});

The agent run only resumes once every open interrupt has been addressed. Calling resolve or cancel with a specific interruptId marks that interrupt done; the hook auto-submits the accumulated responses when the last one is addressed. If you omit interruptId, the primary interrupt (interrupts[0]) is targeted.

responseSchema: surface only, no client-side validation

The Interrupt type exposes a responseSchema field (a JSON Schema object) that the agent can use to describe the expected payload shape. useInterrupt surfaces this field on interrupt.responseSchema for your UI to read (e.g. to drive a form), but it does not validate resolve payloads against it. Validation is the agent's responsibility on resume.

How is this guide?