OpenAI Agents SDK

Run your OpenAI Agents SDK agent for each incoming message and pass its stream to Flow’s reply(). The messaging integration is the same on each connected channel.

TypeScript

// npm install @flow-engineer/messaging @openai/agents
import { Agent, run } from "@openai/agents";
import { FlowMessaging, contentText } from "@flow-engineer/messaging";

const flow = new FlowMessaging();
const agent = new Agent({
  name: "Bakery assistant",
  instructions: "You help customers of Asha's Bakery with orders and delivery. Keep answers short.",
  model: "gpt-4.1-mini",
});

for await (const event of flow.events.stream({ types: ["message.received"] })) {
  const result = await run(agent, contentText(event.data.message.content), { stream: true });
  await event.conversation.reply(result); // reads result.toTextStream()
}

Python (HTTP)

# pip install openai-agents fastapi requests
# Plain HTTP, no Flow SDK; the Python SDK (pip install flow-messaging) wraps the same calls. This calls the HTTP API.
import os, requests
from fastapi import FastAPI, Request, BackgroundTasks
from agents import Agent, Runner

API = "https://api.flow.engineer"
HEADERS = {"Authorization": f"Bearer {os.environ['FLOW_MESSAGING_KEY']}"}
agent = Agent(name="Bakery assistant", instructions="You help customers of Asha's Bakery. Keep answers short.")
app = FastAPI()

async def answer(event: dict):
    text = event["data"]["message"]["content"].get("text", "")
    result = await Runner.run(agent, text)
    requests.post(
        f"{API}/v0/conversations/{event['conversation']['id']}/basic",
        headers={**HEADERS, "Idempotency-Key": event["id"]},
        json={"text": result.final_output},  # sent exactly as written: no markdown is rendered
        timeout=30,
    )

@app.post("/flow")
async def flow_webhook(request: Request, tasks: BackgroundTasks):
    event = await request.json()  # verify Flow-Signature first: see Events and webhooks
    if event["type"] == "message.received":
        tasks.add_task(answer, event)
    return {}

How it fits

• In TypeScript, run(agent, input, { stream: true }) returns a streamed result; reply() reads its text stream and sends bubbles while the agent runs.
• In Python, answer the webhook with {} at once and send the agent’s final_output afterwards. Text longer than the channel’s limit (4096 characters on Telegram) is refused, so split long answers at paragraph breaks (the bubble rule) and send each part with its own Idempotency-Key.
• Give the agent the conversation’s history from GET /v0/conversations/{conversation_id}/messages, or keep your own session keyed by conversation.id.
• Verify webhook signatures as shown in Events and webhooks.
• To see each turn’s model and tool calls and get every reply graded, send this agent’s traces to Flow Observe: Traces from agent frameworks.

Full example

examples/openai-agents on GitHub.