diff --git a/examples/multi-task/local-handoff/local-handoff-two-agents.py b/examples/multi-task/local-handoff/local-handoff-two-agents.py new file mode 100644 index 000000000..21f388d41 --- /dev/null +++ b/examples/multi-task/local-handoff/local-handoff-two-agents.py @@ -0,0 +1,243 @@ +# +# Copyright (c) 2026, Daily +# +# SPDX-License-Identifier: BSD 2-Clause License +# + +"""Two LLM tasks with a main task bridging transport to the bus. + +Demonstrates multi-task coordination: a main task handles transport I/O +(STT, TTS) and bridges frames to the bus. Two LLM tasks — a greeter and +a support task — each run their own LLM pipeline and hand off control +between each other. + +The user talks to one task at a time. Hand-offs are seamless — the LLM +decides when to transfer based on its tools. + +Requirements: + +- OPENAI_API_KEY +- DEEPGRAM_API_KEY +- CARTESIA_API_KEY +- DAILY_API_KEY (for Daily transport) +""" + +import os + +from dotenv import load_dotenv +from loguru import logger + +from pipecat.audio.vad.silero import SileroVADAnalyzer +from pipecat.bus import BusBridgeProcessor +from pipecat.pipeline.pipeline import Pipeline +from pipecat.pipeline.runner import PipelineRunner +from pipecat.pipeline.task import PipelineParams, PipelineTask +from pipecat.processors.aggregators.llm_context import LLMContext +from pipecat.processors.aggregators.llm_response_universal import ( + LLMContextAggregatorPair, + LLMUserAggregatorParams, +) +from pipecat.runner.types import RunnerArguments +from pipecat.runner.utils import create_transport +from pipecat.services.cartesia.tts import CartesiaTTSService +from pipecat.services.deepgram.stt import DeepgramSTTService +from pipecat.services.llm_service import FunctionCallParams +from pipecat.services.openai.llm import OpenAILLMService +from pipecat.tasks.llm import LLMTask, LLMTaskActivationArgs, tool +from pipecat.transports.base_transport import BaseTransport, TransportParams +from pipecat.transports.daily.transport import DailyParams + +load_dotenv(override=True) + +MAIN_NAME = "acme" + + +transport_params = { + "daily": lambda: DailyParams( + audio_in_enabled=True, + audio_out_enabled=True, + ), + "webrtc": lambda: TransportParams( + audio_in_enabled=True, + audio_out_enabled=True, + ), +} + + +class AcmeLLMTask(LLMTask): + """LLM-only child task with transfer/end tools. + + Receives user context from the main task via the bus, runs its LLM, + and ships generated text frames back. The main task's TTS turns the + text into audio. + + Passing ``bridged=()`` tells :class:`PipelineTask` to wrap the LLM + pipeline with bus edge processors so frames flow between this task + and the main task automatically. + """ + + @tool(cancel_on_interruption=False) + async def transfer_to_agent(self, params: FunctionCallParams, agent: str, reason: str): + """Transfer the user to another agent. + + Args: + agent (str): The agent to transfer to (e.g. 'greeter', 'support'). + reason (str): Why the user is being transferred. + """ + logger.info(f"Task '{self.name}': transferring to '{agent}' ({reason})") + await self.handoff_to( + agent, + activation_args=LLMTaskActivationArgs( + messages=[{"role": "developer", "content": reason}], + ), + result_callback=params.result_callback, + ) + + @tool + async def end_conversation(self, params: FunctionCallParams, reason: str): + """End the conversation when the user says goodbye. + + Args: + reason (str): Why the conversation is ending. + """ + logger.info(f"Task '{self.name}': ending conversation ({reason})") + await self.end( + reason=reason, + messages=[{"role": "developer", "content": reason}], + result_callback=params.result_callback, + ) + + +def _build_greeter() -> AcmeLLMTask: + """Greeter: routes the user to support when they pick a product.""" + llm = OpenAILLMService( + api_key=os.environ["OPENAI_API_KEY"], + settings=OpenAILLMService.Settings( + system_instruction=( + "You are a friendly greeter for Acme Corp. The available products " + "are: the Acme Rocket Boots, the Acme Invisible Paint, and the Acme " + "Tornado Kit. Ask which one they'd like to learn more about. " + "When the user picks a product or asks a question about one, " + "immediately call the transfer_to_agent tool with agent 'support'. " + "Do not answer product questions yourself. If the user says goodbye, " + "call the end_conversation tool. Do not mention transferring — just do it " + "seamlessly. Keep responses brief — this is a voice conversation." + ), + ), + ) + return AcmeLLMTask("greeter", llm=llm, bridged=()) + + +def _build_support() -> AcmeLLMTask: + """Support: answers product questions, can hand back to the greeter.""" + llm = OpenAILLMService( + api_key=os.environ["OPENAI_API_KEY"], + settings=OpenAILLMService.Settings( + system_instruction=( + "You are a support agent for Acme Corp. You know about three " + "products: Acme Rocket Boots (jet-powered boots, $299, run up " + "to 60 mph), Acme Invisible Paint (makes anything invisible for " + "24 hours, $49 per can), and Acme Tornado Kit (portable tornado " + "generator, $199, batteries included). Answer the user's questions " + "about these products. If the user wants to browse other products " + "or start over, call the transfer_to_agent tool with agent " + "'greeter'. If the user says goodbye, call the end_conversation " + "tool. Do not mention transferring — just do it seamlessly. " + "Keep responses brief — this is a voice conversation." + ), + ), + ) + return AcmeLLMTask("support", llm=llm, bridged=()) + + +async def run_bot(transport: BaseTransport, runner_args: RunnerArguments): + logger.info("Starting two-agent bot") + + runner = PipelineRunner(handle_sigint=runner_args.handle_sigint) + + stt = DeepgramSTTService(api_key=os.environ["DEEPGRAM_API_KEY"]) + tts = CartesiaTTSService( + api_key=os.environ["CARTESIA_API_KEY"], + settings=CartesiaTTSService.Settings( + voice="9626c31c-bec5-4cca-baa8-f8ba9e84c8bc", # Jacqueline + ), + ) + + context = LLMContext() + aggregators = LLMContextAggregatorPair( + context, + user_params=LLMUserAggregatorParams(vad_analyzer=SileroVADAnalyzer()), + ) + + # The main bridge sends user-side context downstream to the children + # via the bus, and the children's generated text comes back here so + # the TTS can speak it. + bridge = BusBridgeProcessor( + bus=runner.bus, + task_name=MAIN_NAME, + name=f"{MAIN_NAME}::BusBridge", + ) + + pipeline = Pipeline( + [ + transport.input(), + stt, + aggregators.user(), + bridge, + tts, + transport.output(), + aggregators.assistant(), + ] + ) + + task = PipelineTask( + pipeline, + name=MAIN_NAME, + params=PipelineParams( + enable_metrics=True, + enable_usage_metrics=True, + ), + idle_timeout_secs=runner_args.pipeline_idle_timeout_secs, + ) + + # Spawn the child LLM tasks. ``bridged=()`` on each child auto-wraps + # its pipeline with bus edges, so no extra wiring is needed here. + await runner.spawn(_build_greeter()) + await runner.spawn(_build_support()) + + @transport.event_handler("on_client_connected") + async def on_client_connected(transport, client): + logger.info("Client connected") + await task.activate_task( + "greeter", + args=LLMTaskActivationArgs( + messages=[ + { + "role": "developer", + "content": ( + "Welcome the user to Acme Corp, mention the available products " + "and ask how you can help." + ), + }, + ], + ), + ) + + @transport.event_handler("on_client_disconnected") + async def on_client_disconnected(transport, client): + logger.info("Client disconnected") + await task.cancel() + + await runner.run(task) + + +async def bot(runner_args: RunnerArguments): + """Main bot entry point compatible with Pipecat Cloud.""" + transport = await create_transport(runner_args, transport_params) + await run_bot(transport, runner_args) + + +if __name__ == "__main__": + from pipecat.runner.run import main + + main()