agent-wait

Get a LangGraph interrupt out of the process, and the answer back in.

When a node calls interrupt(), 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 a person on 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 pause and puts it somewhere people can see it — a topic, a queue, a webhook, a database row — with everything needed to answer it in one envelope, and reduces "is this message a new request or an answer?" to a one-line check.

It does not receive the answer for you. That part is yours, and it is about a dozen lines.

pip install agent-wait langgraph-wait          # core + LangGraph
pip install agent-wait-aws                     # SNS / SQS / EventBridge / DynamoDB announcers

The whole thing

In the graph — one line, where the decision belongs:

from agent_wait import WaitPolicy
from langgraph_wait import ask


def review(state):
    if state["amount"] <= 5_000:
        return {"decision": {"action": "approve", "by": "policy:auto"}}

    decision = ask(
        {"kind": "refund_approval", "order_id": state["order_id"], "amount": state["amount"]},
        policy=WaitPolicy(
            timeout="P3D",
            default={"action": "reject", "reason": "no response in 3 days"},
            allowed_actions=("approve", "reject"),
            tags={"approver_group": "finance"},
        ),
    )
    return {"decision": decision}

ask() is a thin wrapper over interrupt(). The node pauses exactly as LangGraph pauses; what ask() adds is the policy, which rides along and comes back out in the envelope. A plain interrupt(value) works too, with default policy — a graph that already interrupts gets published with no edit at all.

In the host — wire it once:

from agent_wait import WaitPublisher
from agent_wait_aws import SnsAnnounce
from langgraph_wait import LangGraphAdapter

agent = WaitPublisher(LangGraphAdapter(graph), announce=[SnsAnnounce(topic_arn)])

Then route each message. Starts and answers arrive at the same place; interrupt_id tells them apart:

from langgraph_wait import is_answer, resume_command


def route(message):
    thread_id = message["thread_id"]
    if is_answer(message):
        if not is_still_open(thread_id, message["interrupt_id"]):
            return  # somebody already answered
        return agent.invoke(resume_command(message), thread_id)
    if agent.pending(thread_id):
        return agent.republish(thread_id)  # a redelivery; don't re-ask
    return agent.invoke(message["input"], thread_id)


def is_still_open(thread_id, interrupt_id):
    return any(p.interrupt_id == interrupt_id for p in agent.pending(thread_id))

That is the complete integration. examples/refund_agent/ is it, deployed to Lambda behind SQS.

What goes out

{
  "type": "wait.created",
  "thread_id": "order-4471",
  "interrupt_id": "a1b2c3d4e5f60718",
  "question": { "kind": "refund_approval", "amount": 41000 },
  "allowed_actions": ["approve", "reject"],
  "expires_at": "2026-09-12T09:00:00Z",
  "default": { "action": "reject", "reason": "no response in 3 days" },
  "reply_with": { "thread_id": "order-4471", "interrupt_id": "a1b2c3d4e5f60718", "answer": null }
}

reply_with is a filled-in stub: the consumer copies it, sets answer, and posts it to wherever your agent listens. Whatever goes in answer is what the ask() call returns — verbatim, with nothing merged into it.

A second envelope, wait.resumed, goes out when the graph moves past the question, so a UI knows to retract the button.

Full schema, including how to deduplicate: Message formats.

Announcers

An announcer is the only thing you are expected to implement. Subclass BaseAnnounce and 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())

The contract — an announcer must never raise into the run — is enforced by the base class: an exception from deliver() becomes a log line, and the graph that just parked stays parked.

Because nothing reads state back through this library, "announce" doesn't have to mean "publish an event". It means put the question where whoever answers it will find it:

Pass as many as you like; failures are contained per adapter. The full guide — what deliver() receives, why dedupe_key is the one field to get right, patterns from the shipped adapters, and how to test yours: Writing an announcer.

What the library does not do

Deliberately — each of these is where teams' own opinions live:

• Receive answers. No inbound endpoint, no validation, no tokens. The router above is yours.
• Enforce the timeout. expires_at and default are published; a sweep of yours sends the default when the deadline passes. There is a working one in the example.
• Decide a race. pending() rejects an answer the graph has already moved past. Two different answers in the same instant are your transport's problem — SQS FIFO keyed by thread solves it; an HTTP endpoint with concurrent handlers needs a conditional write.
• Authenticate. Whoever can write to your entry point can answer.
• Store anything. LangGraph's checkpoint is the only state.

Two LangGraph 1.2.x behaviours you should know about

Both verified against 1.2.11, both pinned by tests that fail if LangGraph changes them.

get_state().tasks[*].interrupts over-reports (#4796, #6792). Resume one of two parallel interrupts and the finished task still lists its id. pending() filters on task.result, which is None only while genuinely parked.

Two interrupting tools in one ToolNode get the same id (#6626, #6624). A different question under an identical id defeats deduplication, and there is no filter for it. The rule is one interrupt() per node — give each approval-requiring tool its own node, which is also the fix for a node re-running its side effects on resume.

Details: Architecture.

Layout

packages/agent-wait        core. No LangGraph, no AWS, no dependencies. pyright strict.
packages/langgraph-wait    ask(), the adapter, resume_command(). The only LangGraph import.
packages/agent-wait-aws    four announce adapters, and a CDK stack.
examples/refund_agent      a graph, a router, and four scenarios against real AWS.
docs/                      message contract, architecture, announcer guide, consumer guide.

uv sync
uv run pytest
uv run ruff check . && uv run pyright

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