agent-wait

Make an agent wait for something outside the process — and get the answer back in.

Human-in-the-loop is the common case: a refund over the limit, a contract clause, a deploy. But the same shape covers a vendor callback, a payment processor, a KYC check, or another agent — anything the run cannot continue without, and nothing inside the process can supply.

LangGraph has interrupt() for this. When a node calls it the graph pauses and the interrupt is handed to whatever called invoke() — and that is where LangGraph stops. There is no built-in way to tell anyone else that a question was asked, and no built-in way for anyone else to answer it. The moment the question has to reach Slack, an approvals dashboard, a ticket queue or another service, you are writing that code yourself — whether your agent is a server that runs for a year or a Lambda that is gone in seconds.

agent-wait is that code. It takes the interrupt and puts it somewhere people (or systems) can see it — a topic, a queue, a webhook, a database row — with everything needed to answer it in one envelope, and documents the shape of the answer so the return leg is one if.

pip install "agent-wait[langgraph]"        # @wait + publish_interrupts for LangGraph
pip install "agent-wait[langgraph,aws]"    # + SNS / SQS / EventBridge / DynamoDB announcers

The bare agent-wait is the framework-free core (policy, envelope, webhook and in-memory announcers) and depends on nothing. Each extra pulls in exactly one framework or provider.

The whole thing

In the graph — one decorator, on the function where the decision belongs:

from agent_wait import WaitPolicy
from agent_wait.langgraph import wait
from langchain_core.tools import tool

FINANCE = WaitPolicy(
    timeout="P3D",
    default={"action": "reject"},
    allowed_actions=("approve", "reject"),
    tags={"approver_group": "finance"},
)


@tool
@wait(FINANCE)
def issue_refund(order_id: str, amount: int) -> str:
    payments.refund(order_id, amount)  # runs only if the answer is {"action": "approve"}
    return "refunded"

Calling issue_refund(...) inside a run parks the graph on {"function": "issue_refund", "args": {...}}. On resume, {"action": "approve"} runs the body — optionally with edited args — and anything else is returned in its place, so a model sees why it did not run.

The policy is what the asker declares about the question. Every field is published and none is enforced — the consumer acts on them:

A node that needs the answer itself declares a decision parameter and always runs, approve or not, with the answer in it:

@wait(FINANCE)
def review(state, decision=None):
    return {"decision": decision}  # verbatim: {"action": "approve", "note": "ok"}

The parameter name is yours — @wait(FINANCE, decision="verdict") looks for verdict.

After the run — one call:

from agent_wait.aws import SnsAnnounce
from agent_wait.langgraph import publish_interrupts

result = graph.invoke(value, {"configurable": {"thread_id": thread_id}})
publish_interrupts(result, thread_id, announce=[SnsAnnounce(topic_arn)])

It reads result["__interrupt__"] — Interrupt.id and Interrupt.value, nothing else — and hands one envelope per question to the announcers. No graph handle, no get_state(), no checkpointer. If nothing interrupted, it does nothing. A plain interrupt(value) from a node that never heard of this library is published too, with the default policy.

When the answer comes back — the one if you write:

from langgraph.types import Command

if "question_id" in message:  # an answer
    value = Command(resume={message["question_id"]: message["answer"]})
else:  # a start
    value = message["input"]

result = graph.invoke(value, {"configurable": {"thread_id": message["thread_id"]}})
publish_interrupts(result, message["thread_id"], announce)

That is the complete integration. A duplicate answer runs nothing — LangGraph ignores a resume for a question the thread has moved past — so there is nothing to check.

Async mode — the thread does not park

from agent_wait.aws import DynamoDbAnnounce, SnsAnnounce


@tool
@wait(FINANCE, mode="async", announce=[SnsAnnounce(topic_arn), DynamoDbAnnounce(table)])
def issue_refund(order_id: str, amount: int) -> str: ...

The decorator is the publisher: the call announces the question and returns {"status": "pending_approval", "question_id": "...", "function": "issue_refund"} without running the body. The graph carries on; nothing is parked; nothing to call after the run. The decision arrives later as a new message and your graph acts on it. This is the mode for a single-thread channel like WhatsApp, where the approver is not the person on the thread — and since nothing is parked, LangGraph remembers nothing about the question; the DynamoDbAnnounce row (or whatever you keep) is the only record.

What goes out

{
  "type": "wait.created",
  "thread_id": "order-4471",
  "question_id": "a1b2c3d4e5f60718",
  "question": { "function": "issue_refund", "args": { "order_id": "order-4471", "amount": 41000 } },
  "allowed_actions": ["approve", "reject"],
  "expires_at": "2026-09-14T09:00:00Z",
  "default": { "action": "reject" },
  "source": { "function": "issue_refund" },
  "reply_with": { "thread_id": "order-4471", "question_id": "a1b2c3d4e5f60718", "answer": null }
}

reply_with is a filled-in stub: copy it, set answer, send it back. Whatever goes in answer is what the @wait function receives — verbatim.

Full schema, the recommended answer shape, and how to deduplicate: Message formats.

Announcers

Subclass BaseAnnounce, write one method:

from agent_wait import BaseAnnounce


class RedisAnnounce(BaseAnnounce):
    name = "redis"

    def __init__(self, client, **kw):
        super().__init__(**kw)
        self.client = client

    def deliver(self, envelope, transition):
        self.client.set(envelope.dedupe_key, envelope.to_json())

A raise inside deliver() becomes a log line; the run completes.

Shipped in the core: WebhookAnnounce (signed JSON POST, stdlib), LogAnnounce, InMemoryAnnounce. With [aws]: SnsAnnounce, SqsAnnounce, EventBridgeAnnounce, DynamoDbAnnounce (the question as a row an approvals UI can query). Each one, its constructor and where the question lands: Announcers. Your own: Writing an announcer.

What the library does not do

• Receive answers. No endpoint, no validation, no ledger reads. The if above is yours.
• Enforce anything. expires_at, default, answer_ttl, allowed_actions are published so the consumer has the asker's intent. Acting on them is the consumer's.
• Store anything. LangGraph's checkpoint is the only state in interrupt mode; in async mode there is none unless you keep one.

Two things to know

Not compatible with LangChain's HumanInTheLoopMiddleware. It interrupts before a tool is called, in its own shape; @wait interrupts inside the call. Both on one tool means two interrupts for one approval. Use one or the other.

One interrupt() per node on langgraph 1.2.x. Two interrupting tools dispatched by one ToolNode get the same interrupt id (#6626), and only one surfaces per run (#6624). Give each interrupting tool its own node. Pinned by a test that fails if LangGraph changes it.

Layout

One distribution, one import root. Frameworks and providers are subpackages behind extras:

agent_wait              core: WaitPolicy, Question, the envelope, publish(), BaseAnnounce,
                        Webhook/Log/InMemory announcers, and the Framework interface. No deps.
agent_wait.langgraph    [langgraph]  @wait and publish_interrupts() bound to LangGraph.
agent_wait.aws          [aws]        SnsAnnounce, SqsAnnounce, EventBridgeAnnounce, DynamoDbAnnounce.
examples/refund_agent   a graph, a host written for Lambda behind SQS, and its CDK stack.
docs/                   message contract, architecture, announcer guide, consumer guide.

@wait and publish_interrupts are written once against a three-method Framework interface; agent_wait.langgraph is one implementor of it. Announcers are the same split: BaseAnnounce is the interface, each provider subpackage the implementors. A strands extra follows the same pattern — see Architecture.

MIT. Issues and PRs at github.com/skamalj/agent-wait.