실무 가이드: 오픈 소스 에이전트 프레임워크와 ElevenAgents

def sse_chunk(response_id: str, delta: dict, finish_reason=None) -> str:
    payload = {
        "id": response_id,
        "object": "chat.completion.chunk",
        "choices": [{"index": 0, "delta": delta, "finish_reason": finish_reason}],
    }
    return f"data: {json.dumps(payload)}\n\n"

from langchain.agents import create_agent
from langchain_openai import ChatOpenAI
from langchain_core.tools import tool
llm = ChatOpenAI(
    model=model_id,
    api_key=os.getenv("OPENAI_API_KEY"),
)
agent = create_agent(
    llm,
	tools=tool_list,
	system_prompt=system_prompt,
)

@app.post("/chat/completions")
async def chat_completions(req: ChatCompletionRequest):
    input = {"messages": req.messages}
    async def stream():
        response_id = f"chatcmpl-{uuid.uuid4().hex[:12]}"
        sent_role = False
        async for message_chunk, metadata in agent.astream(input, stream_mode="messages"):
            # Only forward model text chunks; skip tool updates and non-text events.
            if metadata.get("langgraph_node") != "model":
                continue
            content = getattr(message_chunk, "content", None)
            if not content:
                continue
            if not sent_role:
                yield sse_chunk(response_id, {"role": "assistant"})
                sent_role = True
            # Send incremental token-like chunks to ElevenLabs in OpenAI format.
            yield sse_chunk(response_id, {"content": content})
         # Signal natural completion before using the finish_reason: "stop" [DONE]
        yield sse_chunk(response_id, {}, finish_reason="stop")
        yield "data: [DONE]\n\n"
    return StreamingResponse(stream(), media_type="text/event-stream")

from google.adk.agents import Agent
from google.adk.runners import Runner
from google.adk.agents.run_config import RunConfig, StreamingMode
from google.adk.sessions import InMemorySessionService
from google.genai import types as genai_types
agent = Agent(
    name=name,
    model=model,
    instruction=instruction,
    tools=[tool_list],
)
session_service = InMemorySessionService()
	runner = Runner(
	agent=agent,
	app_name=app_name,
	session_service=session_service
)

@app.post("/chat/completions")
async def chat_completions(req: ChatCompletionRequest, request: Request):
    # In production, prefer a stable identifier from your upstream system.
    session_id = req.elevenlabs_extra_body.arbitrary_identifier
    session = await session_service.get_session(
        app_name="elevenlabs", user_id="user", session_id=session_id
    )
    if not session:
        session = await session_service.create_session(
            app_name="elevenlabs", user_id="user", session_id=session_id
        )
    user_text = next((m["content"] for m in reversed(req.messages) if m["role"] == "user"), "")
    content = genai_types.Content(role="user", parts=[genai_types.Part(text=user_text)])
   async def stream():
        response_id = f"chatcmpl-{uuid.uuid4().hex[:12]}"
        sent_role = False
        async for event in runner.run_async(
            user_id="user",
            session_id=session.id,
            new_message=content,
            run_config=RunConfig(streaming_mode=StreamingMode.SSE),
        ):
            if not event.content or not event.content.parts:
                continue
            # In SSE mode, ADK emits partial (incremental) and final (complete) events.
            # Forwarding only partial events avoids duplicating the full text.
            # Note: SSE streaming is experimental in ADK. For production, reconcile
            # both event types in case the model backend doesn't emit partials.
            if not getattr(event, "partial", False):
                continue
            text = "".join((getattr(p, "text", "") or "") for p in event.content.parts)
            if not text:
                continue
            if not sent_role:
                yield sse_chunk(response_id, {"role": "assistant"})
                sent_role = True
            yield sse_chunk(response_id, {"content": text})
        yield sse_chunk(response_id, {}, finish_reason="stop")
        yield "data: [DONE]\n\n"
    return StreamingResponse(stream(), media_type="text/event-stream")

from crewai import Agent, Task, Crew, Process, LLM
from crewai.tools import tool
from crewai.types.streaming import StreamChunkType
llm = LLM(
    model=model_id,
    api_key=os.getenv("OPENAI_API_KEY")
)
store_agent = Agent(
    role=role,
    goal=goal,
    backstory=backstory,
    tools=tools,
    llm=llm,
    verbose=False,
)

@app.post("/chat/completions")
async def chat_completions(req: ChatCompletionRequest):
    # Task and Crew are assembled per request (not at startup).	
    task = Task(
        description=(
            "Conversation history:\n{crew_chat_messages}\n\n"
            "Respond to the user's latest message."
        ),
        expected_output=expected_output,
        agent=store_agent,
    )
    # stream=True returns CrewStreamingOutput instead of a single CrewOutput.
    crew = Crew(
        agents=[store_agent],
        tasks=[task],
        process=Process.sequential,
        verbose=False,
        stream=True,
    )
    async def stream():
        response_id = f"chatcmpl-{uuid.uuid4().hex[:12]}"
        sent_role = False
        final_marker = "final answer:"
        marker_buffer = ""
        marker_found = False
        emitted_any_content = False
        streaming = await crew.kickoff_async(
            inputs={"crew_chat_messages": json.dumps(req.messages)}
        )
       async for chunk in streaming:
            # Skip non-text events (e.g. tool calls).
            if chunk.chunk_type != StreamChunkType.TEXT or not chunk.content:
                continue
            # Only forward text after the "Final Answer:" marker
            if not marker_found:
                marker_buffer += chunk.content
                idx = marker_buffer.lower().find("final answer:")
                if idx == -1:
                    continue
                marker_found = True
                content = marker_buffer[idx + 13:].lstrip()
                marker_buffer = ""
            else:
                content = chunk.content
            # Clean up any trailing markdown artifacts from CrewAI output.
            content = content.rstrip("`").rstrip()
            if not content:
                continue
            if not sent_role:
                yield sse_chunk(response_id, {"role": "assistant"})
                sent_role = True
            yield sse_chunk(response_id, {"content": content})
        # Fallback to handle short responses without the "Final Answer:" marker
        if not sent_role:
            raw = getattr(streaming, "result", None)
            fallback = (raw.raw if raw else marker_buffer).strip().rstrip("`").rstrip()
            if fallback:
                yield sse_chunk(response_id, {"role": "assistant"})
                yield sse_chunk(response_id, {"content": fallback})
        yield sse_chunk(response_id, {}, finish_reason="stop")
        yield "data: [DONE]\n\n"

from llama_index.llms.openai import OpenAI
from llama_index.core.agent.workflow import FunctionAgent, AgentStream
from llama_index.core.base.llms.types import ChatMessage, MessageRole
llm = OpenAI(
    model=model,
    api_key=os.getenv("OPENAI_API_KEY")
)
agent = FunctionAgent(
    tools=[list_inventory, get_item_price],
    llm=llm,
    system_prompt=system_prompt,
)

@app.post("/chat/completions")
async def chat_completions(req: ChatCompletionRequest):
    # This assumes the last message is always a user turn with string content.
    # For production, add defensive role/content handling for non-text payloads.
    chat_history = [
        ChatMessage(role=MessageRole(m["role"]), content=m.get("content") or "")
        for m in req.messages
    ]
    user_text = chat_history.pop().content
    async def stream():
        response_id = f"chatcmpl-{uuid.uuid4().hex[:12]}"
        handler = agent.run(user_msg=user_text, chat_history=chat_history)
        async for event in handler.stream_events():
            if not isinstance(event, AgentStream):
                continue
            if not event.delta:
                continue
            yield sse_chunk(response_id, {"content": event.delta})
        yield sse_chunk(response_id, {}, finish_reason="stop")
        yield "data: [DONE]\n\n"
    return StreamingResponse(stream(), media_type="text/event-stream")