Multi-Agent 架構(gòu)之 Supervisor
作者：theanarkh
本文基于開源的 langgraph-supervisor-py 項(xiàng)目，分析 supervisor 架構(gòu)下的多 Agent 設(shè)計(jì)，supervisor 架構(gòu)是有一個(gè) supervisor 和多個(gè) Agent，supervisor 負(fù)責(zé)管理多個(gè) Agent，Agent 負(fù)責(zé)具體的工作，開始執(zhí)行時(shí)，從 supervisor 開始，然后根據(jù)大模型的返回不斷調(diào)各個(gè) Agent，每個(gè) Agent 執(zhí)行完再次回
從單 Agent 到 Workflow，現(xiàn)在已經(jīng)進(jìn)入了多 Agent 時(shí)代，多 Agent 本質(zhì)上是在大模型的指導(dǎo)下，多個(gè) Agent 共同協(xié)作完成用戶的任務(wù)。在這個(gè)背景下衍生出了多種 Agent 架構(gòu)，比如中心化、去中心化、層次化等。本文基于開源的 langgraph-supervisor-py 項(xiàng)目，分析 supervisor 架構(gòu)下的多 Agent 設(shè)計(jì)，supervisor 架構(gòu)是有一個(gè) supervisor 和多個(gè) Agent，supervisor 負(fù)責(zé)管理多個(gè) Agent，Agent 負(fù)責(zé)具體的工作，開始執(zhí)行時(shí)，從 supervisor 開始，然后根據(jù)大模型的返回不斷調(diào)各個(gè) Agent，每個(gè) Agent 執(zhí)行完再次回到 supervisor，supervisor 再次調(diào)大模型，如此反復(fù)，直到得到結(jié)果。
使用例子
下面從官方的一個(gè)例子開始。
from langchain_openai import ChatOpenAI
from langgraph_supervisor import create_supervisor
from langgraph.prebuilt import create_react_agent
from langchain_core.runnables.graph import MermaidDrawMethod
# 工具
def add(a: float, b: float) -> float:
"""Add two numbers."""
return a + b
def multiply(a: float, b: float) -> float:
"""Multiply two numbers."""
return a * b
def web_search(query: str) -> str:
"""Search the web for information."""
return (
"Here are the headcounts for each of the FAANG companies in 2024:\n"
"1. **Facebook (Meta)**: 67,317 employees.\n"
"2. **Apple**: 164,000 employees.\n"
"3. **Amazon**: 1,551,000 employees.\n"
"4. **Netflix**: 14,000 employees.\n"
"5. **Google (Alphabet)**: 181,269 employees."
)
# 數(shù)學(xué)計(jì)算 Agent
math_agent = create_react_agent(
model=model,
tools=[add, multiply],
name="math_expert",
prompt="You are a math expert. Always use one tool at a time."
)
# 搜索 Agent
research_agent = create_react_agent(
model=model,
tools=[web_search],
name="research_expert",
prompt="You are a world class researcher with access to web search. Do not do any math."
)
workflow = create_supervisor(
[research_agent, math_agent],
model=model,
prompt=(
"You are a team supervisor managing a research expert and a math expert. "
"For current events, use research_agent. "
"For math problems, use math_agent."
)
)
app = workflow.compile()
# 畫圖
img = app.get_graph().draw_mermaid_png(
draw_method=MermaidDrawMethod.API,
)
with open("graph.png", "wb") as f:
f.write(img)
f.close()
# 執(zhí)行
result = app.invoke({
"messages": [
{
"role": "user",
"content": "what's the combined headcount of the FAANG companies in 2024?"
}
]
})上面的例子中創(chuàng)建了一個(gè) supervisor 和兩個(gè) Agent（supervisor 其實(shí)也是一個(gè) Agent），一個(gè)是用于搜索的 Agent，一個(gè)是用于做數(shù)學(xué)計(jì)算的 Agent，每個(gè) Agent 擁有自己的工具集。三個(gè)節(jié)點(diǎn)形成以下結(jié)構(gòu)的圖。
接著詳細(xì)分析上面的例子。
create_react_agent
create_react_agent 可以創(chuàng)建一個(gè) react Agent，其工作原理
- Agent 調(diào)用大模型獲取工具調(diào)用情況。
- 大模型返回調(diào)用哪個(gè)工具。
- Agent 調(diào)用工具，拿到結(jié)果，回到 1，除非返回不需要再調(diào)工具。 其實(shí)現(xiàn)如下。
def create_react_agent(
model: Union[str, LanguageModelLike],
tools: Union[Sequence[Union[BaseTool, Callable]], ToolNode],
*,
prompt: Optional[Prompt] = None,
...
) -> CompiledGraph:
# 創(chuàng)建一個(gè) Tool 節(jié)點(diǎn)統(tǒng)一處理 tool
tool_node = ToolNode(tools)
tool_classes = list(tool_node.tools_by_name.values())
tool_calling_enabled = len(tool_classes) > 0
# 綁定到 model，調(diào)用大模型時(shí)把工具信息傳過去
if _should_bind_tools(model, tool_classes) and tool_calling_enabled:
model = cast(BaseChatModel, model).bind_tools(tool_classes)
model_runnable = _get_prompt_runnable(prompt) | model
# Agent 節(jié)點(diǎn)的處理函數(shù)
def call_model(state: StateSchema, config: RunnableConfig) -> StateSchema:
state = _get_model_input_state(state)
response = cast(AIMessage, model_runnable.invoke(state, config))
# add agent name to the AIMessage
response.name = name
return ...
# 執(zhí)行完 Agent 后執(zhí)行該判斷函數(shù)
def should_continue(state: StateSchema) -> Union[str, list]:
messages = _get_state_value(state, "messages")
last_message = messages[-1]
# 不需要執(zhí)行工具調(diào)用則返回結(jié)束
if not isinstance(last_message, AIMessage) or not last_message.tool_calls:
return END if response_format is None else "generate_structured_response"
# 跳到 Tool 節(jié)點(diǎn)處理
else:
if version == "v1":
return "tools"
elif version == "v2":
tool_calls = [
tool_node.inject_tool_args(call, state, store) # type: ignore[arg-type]
for call in last_message.tool_calls
]
return [Send("tools", [tool_call]) for tool_call in tool_calls]
# 開始構(gòu)建執(zhí)行流程圖
workflow = StateGraph(state_schema or AgentState, config_schema=config_schema)
# 添加 Agent 節(jié)點(diǎn)
workflow.add_node(
"agent", RunnableCallable(call_model, acall_model), input=input_schema
)
# 添加 Tool 節(jié)點(diǎn)
workflow.add_node("tools", tool_node)
# 設(shè)置開始執(zhí)行節(jié)點(diǎn)
entrypoint = "agent"
workflow.set_entry_point(entrypoint)
# 設(shè)置條件判斷節(jié)點(diǎn)
workflow.add_conditional_edges(
"agent",
should_continue,
path_map=should_continue_destinations,
)
# 設(shè)置 Tool 節(jié)點(diǎn)到 Agent 的表，Tool 執(zhí)行完后自動(dòng)跳到 Agent 節(jié)點(diǎn)
workflow.add_edge("tools", entrypoint)
return workflow.compile(
checkpointer=checkpointer,
store=store,
interrupt_before=interrupt_before,
interrupt_after=interrupt_after,
debug=debug,
name=name,
)create_react_agent 內(nèi)部是使用 langgraph 構(gòu)建了一個(gè)執(zhí)行圖，并封裝了大模型和工具的調(diào)用邏輯。當(dāng)開始執(zhí)行 Agent 時(shí)，會(huì)從圖起點(diǎn)開始執(zhí)行，也就是調(diào)用大模型拿到需要調(diào)用的工具（也可能不需要執(zhí)行工具，這里需要?jiǎng)討B(tài)判斷），然后執(zhí)行到 ToolNode 節(jié)點(diǎn)，ToolNode 節(jié)點(diǎn)內(nèi)部會(huì)根據(jù)大模型的返回信息調(diào)用具體的工具，調(diào)用完工具后回到 Agent，Agent 再次調(diào)大模型，如此反復(fù)，工具的處理函數(shù)如下。
def _run_one(
self,
call: ToolCall,
input_type: Literal["list", "dict", "tool_calls"],
config: RunnableConfig,
) -> ToolMessage:
input = {**call, **{"type": "tool_call"}}
response = self.tools_by_name[call["name"]].invoke(input, config)
if isinstance(response, Command):
return self._validate_tool_command(response, call, input_type)
elif isinstance(response, ToolMessage):
response.content = cast(
Union[str, list], msg_content_output(response.content)
)
return response
else:
raise TypeError(
f"Tool {call['name']} returned unexpected type: {type(response)}"
)可以看到 ToolNode 內(nèi)部會(huì)根據(jù)工具名稱調(diào)用具體的工具，拿到結(jié)果后再跳回 Agent。結(jié)構(gòu)圖如下。
create_supervisor
接著從 create_supervisor 開始分析 supervisor 內(nèi)部的架構(gòu)。
def create_supervisor(
agents: list[Pregel],
*,
model: LanguageModelLike,
tools: list[BaseTool | Callable] | None = None,
prompt: Prompt | None = None,
state_schema: StateSchemaType = AgentState,
config_schema: Type[Any] | None = None,
add_handoff_messages: bool = True,
handoff_tool_prefix: Optional[str] = None,
add_handoff_back_messages: Optional[bool] = None,
supervisor_name: str = "supervisor",
) -> StateGraph:
# 默認(rèn)把 handoff 信息記錄到 messages 中
if add_handoff_back_messages is None:
add_handoff_back_messages = add_handoff_messages
# 創(chuàng)建 handoff 工具，用于轉(zhuǎn)移處理權(quán)
all_tools = [
create_handoff_tool(
agent_name=agent.name,
name=(None),
add_handoff_messages=add_handoff_messages,
)
for agent in agents
]
# 綁定工具，調(diào)用大模型時(shí)會(huì)傳過去
model = model.bind_tools(all_tools)
# 創(chuàng)建一個(gè)角色為 supervisor 的 Agent
supervisor_agent = create_react_agent(
name=supervisor_name,
model=model,
tools=all_tools,
prompt=prompt,
state_schema=state_schema, # AgentState， reducer 為 add_messages
response_format=response_format,
)
builder = StateGraph(state_schema, config_schema=config_schema)
# 添加 supervisor 到圖中，下游是各個(gè) Agent 和 END
builder.add_node(supervisor_agent, destinatinotallow=tuple(agent_names) + (END,))
# 設(shè)置執(zhí)行起點(diǎn)是 supervisor
builder.add_edge(START, supervisor_agent.name)
for agent in agents:
# 把每個(gè) Agent 加入圖中
builder.add_node(
agent.name,
_make_call_agent(
agent,
output_mode,
add_handoff_back_messages=add_handoff_back_messages,
supervisor_name=supervisor_name,
),
)
# 每個(gè) Agent 都有一條指向 supervisor 的邊
builder.add_edge(agent.name, supervisor_agent.name)
return buildercreate_supervisor 的大概邏輯是創(chuàng)建了一個(gè) supervisor Agent，然后構(gòu)造一個(gè)執(zhí)行圖。有兩個(gè)地方需要分析。分別是 create_handoff_tool 和 _make_call_agent。先看 create_handoff_tool。
def create_handoff_tool(
*,
agent_name: str,
name: str | None = None,
description: str | None = None,
add_handoff_messages: bool = True,
) -> BaseTool:
# 工具的名稱和描述
if name is None:
name = f"transfer_to_{_normalize_agent_name(agent_name)}"
if description is None:
description = f"Ask agent '{agent_name}' for help"
@tool(name, descriptinotallow=description)
def handoff_to_agent(
state: Annotated[dict, InjectedState],
tool_call_id: Annotated[str, InjectedToolCallId],
):
tool_message = ToolMessage(
cnotallow=f"Successfully transferred to {agent_name}",
name=name,
tool_call_id=tool_call_id,
response_metadata={METADATA_KEY_HANDOFF_DESTINATION: agent_name},
)
# 追加 handoff message
handoff_messages = state["messages"] + [tool_message]
# 跳到具體的 Agent
return Command(
goto=agent_name,
graph=Command.PARENT,
update={**state, "messages": handoff_messages},
)
handoff_to_agent.metadata = {METADATA_KEY_HANDOFF_DESTINATION: agent_name}
return handoff_to_agentcreate_handoff_tool 用于創(chuàng)建一個(gè)工具，這些工具信息會(huì)發(fā)送給大模型，大模型會(huì)根據(jù)用戶的問題和工具信息（名稱和描述），并告訴 supervisor 應(yīng)該執(zhí)行哪一個(gè)工具，supervisor 拿到大模型的響應(yīng)后，會(huì)執(zhí)行 create_handoff_tool 函數(shù)，并最終跳轉(zhuǎn)到具體的 Agent。Agent 的執(zhí)行由 _make_call_agent 進(jìn)行了封裝。
def _make_call_agent(
agent: Pregel,
output_mode: OutputMode,
add_handoff_back_messages: bool,
supervisor_name: str,
) -> Callable[[dict], dict] | RunnableCallable:
def _process_output(output: dict) -> dict:
messages = output["messages"]
# 追加 handoff message，這里是追加 handoff 到 supervisor 的信息
if add_handoff_back_messages:
messages.extend(create_handoff_back_messages(agent.name, supervisor_name))
# Agent 有指向 supervisor 的邊，執(zhí)行完后會(huì)自動(dòng)跳到 supervisor
return {
**output,
"messages": messages,
}
def call_agent(state: dict, config: RunnableConfig) -> dict:
# 調(diào)大模型
output = agent.invoke(state, config)
return _process_output(output)
return RunnableCallable(call_agent, ...)_make_call_agent 里做了一些處理并追加 handleoff message 后最終跳轉(zhuǎn)到 supervisor，因?yàn)?Agent 有指向 supervisor 的邊，所以會(huì)自動(dòng)跳轉(zhuǎn)，而不像 supervisor 跳到 Agent 時(shí)需要手動(dòng)控制。從上面的分析可以知道，supervisor 架構(gòu)中，處理流程如下。
- supervisor 傳入 messages 和工具列表，然后調(diào)用大模型，大模型返回調(diào)用哪個(gè)工具。
- handoff 到具體的 Agent。
- Agent 傳入 messages 和自己的工具列表，然后調(diào)用大模型，大模型返回調(diào)用哪個(gè)工具。
- Agent 調(diào)用工具，拿到工具的返回，再次調(diào)用大模型，拿到最終結(jié)果。
- handoff 到 supervisor，回到 1，直到調(diào)用大模型時(shí)，直到返回最終結(jié)果。
調(diào)用日志
下面是前面例子調(diào)用大模型的日志。
- supervisor
# supervisor 調(diào)用大模型，并傳入工具列表，即有哪些 Agent
request: {
"tools": [
{
"type": "function",
"function": {
"name": "transfer_to_research_expert",
"description": "Ask agent 'research_expert' for help",
"parameters": {
"properties": {},
"type": "object"
}
}
},
{
"type": "function",
"function": {
"name": "transfer_to_math_expert",
"description": "Ask agent 'math_expert' for help",
"parameters": {
"properties": {},
"type": "object"
}
}
}
],
"parallel_tool_calls": false,
"messages": [
{
"content": "You are a team supervisor managing a research expert and a math expert. For current events, use research_agent. For math problems, use math_agent.",
"role": "system"
},
{
"content": "what's the combined headcount of the FAANG companies in 2024?",
"role": "user"
}
]
}
# 大模型返回調(diào)用 transfer_to_research_expert Agent
response: {
"id": "02174685630333191645265e627e21d5dcc187f9552fedf6fda0b",
"choices": [
{
"finish_reason": "tool_calls",
"index": 0,
"logprobs": null,
"message": {
"content": "",
"role": "assistant",
"tool_calls": [
{
"id": "call_fyew4c06vn9ide7nhk6sijws",
"function": {
"arguments": " {}",
"name": "transfer_to_research_expert"
},
"type": "function"
}
]
}
}
],
}
}2.handoff 到搜索 Agent
# 搜索 Agent 調(diào)用大模型，并傳入工具列表
request: {
"tools": [
{
"type": "function",
"function": {
"name": "web_search",
"description": "Search the web for information.",
"parameters": {
"properties": {
"query": {
"type": "string"
}
},
"required": [
"query"
],
"type": "object"
}
}
}
],
"messages": [
{
"content": "You are a world class researcher with access to web search. Do not do any math.",
"role": "system"
},
{
"content": "what's the combined headcount of the FAANG companies in 2024?",
"role": "user"
},
{
"content": null,
"name": "supervisor",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_fyew4c06vn9ide7nhk6sijws",
"function": {
"name": "transfer_to_research_expert",
"arguments": "{}"
}
}
]
},
{
"content": "Successfully transferred to research_expert",
"role": "tool",
"tool_call_id": "call_fyew4c06vn9ide7nhk6sijws"
}
]
}
# 大模型返回調(diào)用搜索工具，并提供了參數(shù)
response: {
"id": "02174685630508391645265e627e21d5dcc187f9552fedf724cf2",
"choices": [
{
"finish_reason": "tool_calls",
"index": 0,
"logprobs": null,
"message": {
"content": "",
"role": "assistant",
"tool_calls": [
{
"id": "call_t52sv5g0zm4x1tft5hcnwjjs",
"function": {
"arguments": " {\n \"query\": \"Combined headcount of FAANG companies in 2024\"\n }\n",
"name": "web_search"
},
"type": "function"
}
]
}
}
],
}
# 搜索 Agent 拿到了搜索工具的結(jié)果，并再次調(diào)用大模型
request: {
"tools": [
{
"type": "function",
"function": {
"name": "web_search",
"description": "Search the web for information.",
"parameters": {
"properties": {
"query": {
"type": "string"
}
},
"required": [
"query"
],
"type": "object"
}
}
}
],
"messages": [
{
"content": "You are a world class researcher with access to web search. Do not do any math.",
"role": "system"
},
{
"content": "what's the combined headcount of the FAANG companies in 2024?",
"role": "user"
},
{
"content": null,
"name": "supervisor",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_fyew4c06vn9ide7nhk6sijws",
"function": {
"name": "transfer_to_research_expert",
"arguments": "{}"
}
}
]
},
{
"content": "Successfully transferred to research_expert",
"role": "tool",
"tool_call_id": "call_fyew4c06vn9ide7nhk6sijws"
},
{
"content": null,
"name": "research_expert",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_t52sv5g0zm4x1tft5hcnwjjs",
"function": {
"name": "web_search",
"arguments": "{\"query\": \"Combined headcount of FAANG companies in 2024\"}"
}
}
]
},
{
"content": "Here are the headcounts for each of the FAANG companies in 2024:\n1. **Facebook (Meta)**: 67,317 employees.\n2. **Apple**: 164,000 employees.\n3. **Amazon**: 1,551,000 employees.\n4. **Netflix**: 14,000 employees.\n5. **Google (Alphabet)**: 181,269 employees.",
"role": "tool",
"tool_call_id": "call_t52sv5g0zm4x1tft5hcnwjjs"
}
]
}
# 大模型覺得這個(gè)信息夠了，不需要再搜索
response: {
"id": "02174685630701191645265e627e21d5dcc187f9552fedf6f3f21",
"choices": [
{
"finish_reason": "stop",
"index": 0,
"logprobs": null,
"message": {
"content": "The individual headcounts of FAANG companies in 2024 are as follows: Meta has 67,317 employees, Apple has 164,000 employees, Amazon has 1,551,000 employees, Netflix has 14,000 employees, and Alphabet has 181,269 employees. You can sum these figures to get the combined head - count. ",
"role": "assistant"
}
}
],
}3.handoff 到 supervisor
# handoff 回 supervisor，supervisor 再次調(diào)用大模型
request: {
"tools": [
{
"type": "function",
"function": {
"name": "transfer_to_research_expert",
"description": "Ask agent 'research_expert' for help",
"parameters": {
"properties": {},
"type": "object"
}
}
},
{
"type": "function",
"function": {
"name": "transfer_to_math_expert",
"description": "Ask agent 'math_expert' for help",
"parameters": {
"properties": {},
"type": "object"
}
}
}
],
"parallel_tool_calls": false,
"messages": [
{
"content": "You are a team supervisor managing a research expert and a math expert. For current events, use research_agent. For math problems, use math_agent.",
"role": "system"
},
{
"content": "what's the combined headcount of the FAANG companies in 2024?",
"role": "user"
},
{
"content": null,
"name": "supervisor",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_fyew4c06vn9ide7nhk6sijws",
"function": {
"name": "transfer_to_research_expert",
"arguments": "{}"
}
}
]
},
{
"content": "Successfully transferred to research_expert",
"role": "tool",
"tool_call_id": "call_fyew4c06vn9ide7nhk6sijws"
},
{
"content": "The individual headcounts of FAANG companies in 2024 are as follows: Meta has 67,317 employees, Apple has 164,000 employees, Amazon has 1,551,000 employees, Netflix has 14,000 employees, and Alphabet has 181,269 employees. You can sum these figures to get the combined head - count. ",
"name": "research_expert",
"role": "assistant"
},
{
"content": "Transferring back to supervisor",
"name": "research_expert",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "267ff0b0-d27b-4764-a68b-71eda788379a",
"function": {
"name": "transfer_back_to_supervisor",
"arguments": "{}"
}
}
]
},
{
"content": "Successfully transferred back to supervisor",
"role": "tool",
"tool_call_id": "267ff0b0-d27b-4764-a68b-71eda788379a"
}
]
}
# 大模型返回調(diào)用數(shù)學(xué)計(jì)算 Agent
response: {
"id": "02174685631050391645265e627e21d5dcc187f9552fedf85877e",
"choices": [
{
"finish_reason": "tool_calls",
"index": 0,
"logprobs": null,
"message": {
"content": "",
"role": "assistant",
"tool_calls": [
{
"id": "call_z9x3nnqjvnp9hxuehejkivgn",
"function": {
"arguments": " {}",
"name": "transfer_to_math_expert"
},
"type": "function"
}
]
}
}
],
}4.handoff 到數(shù)學(xué)計(jì)算 Agent
# 數(shù)學(xué)計(jì)算 Agent 調(diào)用大模型，并傳入工具列表
request: {
"tools": [
{
"type": "function",
"function": {
"name": "add",
"description": "Add two numbers.",
"parameters": {
"properties": {
"a": {
"type": "number"
},
"b": {
"type": "number"
}
},
"required": [
"a",
"b"
],
"type": "object"
}
}
},
{
"type": "function",
"function": {
"name": "multiply",
"description": "Multiply two numbers.",
"parameters": {
"properties": {
"a": {
"type": "number"
},
"b": {
"type": "number"
}
},
"required": [
"a",
"b"
],
"type": "object"
}
}
}
],
"messages": [
{
"content": "You are a math expert. Always use one tool at a time.",
"role": "system"
},
{
"content": "what's the combined headcount of the FAANG companies in 2024?",
"role": "user"
},
{
"content": null,
"name": "supervisor",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_fyew4c06vn9ide7nhk6sijws",
"function": {
"name": "transfer_to_research_expert",
"arguments": "{}"
}
}
]
},
{
"content": "Successfully transferred to research_expert",
"role": "tool",
"tool_call_id": "call_fyew4c06vn9ide7nhk6sijws"
},
{
"content": "The individual headcounts of FAANG companies in 2024 are as follows: Meta has 67,317 employees, Apple has 164,000 employees, Amazon has 1,551,000 employees, Netflix has 14,000 employees, and Alphabet has 181,269 employees. You can sum these figures to get the combined head - count. ",
"name": "research_expert",
"role": "assistant"
},
{
"content": "Transferring back to supervisor",
"name": "research_expert",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "267ff0b0-d27b-4764-a68b-71eda788379a",
"function": {
"name": "transfer_back_to_supervisor",
"arguments": "{}"
}
}
]
},
{
"content": "Successfully transferred back to supervisor",
"role": "tool",
"tool_call_id": "267ff0b0-d27b-4764-a68b-71eda788379a"
},
{
"content": null,
"name": "supervisor",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_z9x3nnqjvnp9hxuehejkivgn",
"function": {
"name": "transfer_to_math_expert",
"arguments": "{}"
}
}
]
},
{
"content": "Successfully transferred to math_expert",
"role": "tool",
"tool_call_id": "call_z9x3nnqjvnp9hxuehejkivgn"
}
]
}
# 大模型返回調(diào)用數(shù)學(xué)計(jì)算工具
response: {
"id": "02174685631226191645265e627e21d5dcc187f9552fedf93e39a",
"choices": [
{
"finish_reason": "tool_calls",
"index": 0,
"logprobs": null,
"message": {
"content": "",
"role": "assistant",
"tool_calls": [
{
"id": "call_jd5esfvy2qsh9tdz23361pha",
"function": {
"arguments": " {\n \"a\": 67317, \n \"b\": 164000\n }\n",
"name": "add"
},
"type": "function"
}
]
}
}
],
}
# 數(shù)學(xué)計(jì)算 Agent 調(diào)用大模型，并傳入工具列表
request: {
"tools": [
{
"type": "function",
"function": {
"name": "add",
"description": "Add two numbers.",
"parameters": {
"properties": {
"a": {
"type": "number"
},
"b": {
"type": "number"
}
},
"required": [
"a",
"b"
],
"type": "object"
}
}
},
{
"type": "function",
"function": {
"name": "multiply",
"description": "Multiply two numbers.",
"parameters": {
"properties": {
"a": {
"type": "number"
},
"b": {
"type": "number"
}
},
"required": [
"a",
"b"
],
"type": "object"
}
}
}
],
"messages": [
{
"content": "You are a math expert. Always use one tool at a time.",
"role": "system"
},
{
"content": "what's the combined headcount of the FAANG companies in 2024?",
"role": "user"
},
{
"content": null,
"name": "supervisor",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_fyew4c06vn9ide7nhk6sijws",
"function": {
"name": "transfer_to_research_expert",
"arguments": "{}"
}
}
]
},
{
"content": "Successfully transferred to research_expert",
"role": "tool",
"tool_call_id": "call_fyew4c06vn9ide7nhk6sijws"
},
{
"content": "The individual headcounts of FAANG companies in 2024 are as follows: Meta has 67,317 employees, Apple has 164,000 employees, Amazon has 1,551,000 employees, Netflix has 14,000 employees, and Alphabet has 181,269 employees. You can sum these figures to get the combined head - count. ",
"name": "research_expert",
"role": "assistant"
},
{
"content": "Transferring back to supervisor",
"name": "research_expert",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "267ff0b0-d27b-4764-a68b-71eda788379a",
"function": {
"name": "transfer_back_to_supervisor",
"arguments": "{}"
}
}
]
},
{
"content": "Successfully transferred back to supervisor",
"role": "tool",
"tool_call_id": "267ff0b0-d27b-4764-a68b-71eda788379a"
},
{
"content": null,
"name": "supervisor",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_z9x3nnqjvnp9hxuehejkivgn",
"function": {
"name": "transfer_to_math_expert",
"arguments": "{}"
}
}
]
},
{
"content": "Successfully transferred to math_expert",
"role": "tool",
"tool_call_id": "call_z9x3nnqjvnp9hxuehejkivgn"
},
{
"content": null,
"name": "math_expert",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_jd5esfvy2qsh9tdz23361pha",
"function": {
"name": "add",
"arguments": "{\"a\": 67317, \"b\": 164000}"
}
}
]
},
{
"content": "231317.0",
"role": "tool",
"tool_call_id": "call_jd5esfvy2qsh9tdz23361pha"
}
]
}
# 大模型返回了調(diào)用加法工具，并提供了參數(shù)
response: {
"id": "02174685631455691645265e627e21d5dcc187f9552fedf732682",
"choices": [
{
"finish_reason": "tool_calls",
"index": 0,
"logprobs": null,
"message": {
"content": "",
"role": "assistant",
"tool_calls": [
{
"id": "call_owov4sx70lmdpz7zooldnue3",
"function": {
"arguments": " {\"a\": 231317, \"b\": 1551000}",
"name": "add"
},
"type": "function"
}
]
}
}
],
}
# 數(shù)學(xué)計(jì)算 Agent 拿到加法工具的調(diào)用結(jié)果后，再次調(diào)大模型
request: {
"tools": [
{
"type": "function",
"function": {
"name": "add",
"description": "Add two numbers.",
"parameters": {
"properties": {
"a": {
"type": "number"
},
"b": {
"type": "number"
}
},
"required": [
"a",
"b"
],
"type": "object"
}
}
},
{
"type": "function",
"function": {
"name": "multiply",
"description": "Multiply two numbers.",
"parameters": {
"properties": {
"a": {
"type": "number"
},
"b": {
"type": "number"
}
},
"required": [
"a",
"b"
],
"type": "object"
}
}
}
],
"messages": [
{
"content": "You are a math expert. Always use one tool at a time.",
"role": "system"
},
{
"content": "what's the combined headcount of the FAANG companies in 2024?",
"role": "user"
},
{
"content": null,
"name": "supervisor",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_fyew4c06vn9ide7nhk6sijws",
"function": {
"name": "transfer_to_research_expert",
"arguments": "{}"
}
}
]
},
{
"content": "Successfully transferred to research_expert",
"role": "tool",
"tool_call_id": "call_fyew4c06vn9ide7nhk6sijws"
},
{
"content": "The individual headcounts of FAANG companies in 2024 are as follows: Meta has 67,317 employees, Apple has 164,000 employees, Amazon has 1,551,000 employees, Netflix has 14,000 employees, and Alphabet has 181,269 employees. You can sum these figures to get the combined head - count. ",
"name": "research_expert",
"role": "assistant"
},
{
"content": "Transferring back to supervisor",
"name": "research_expert",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "267ff0b0-d27b-4764-a68b-71eda788379a",
"function": {
"name": "transfer_back_to_supervisor",
"arguments": "{}"
}
}
]
},
{
"content": "Successfully transferred back to supervisor",
"role": "tool",
"tool_call_id": "267ff0b0-d27b-4764-a68b-71eda788379a"
},
{
"content": null,
"name": "supervisor",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_z9x3nnqjvnp9hxuehejkivgn",
"function": {
"name": "transfer_to_math_expert",
"arguments": "{}"
}
}
]
},
{
"content": "Successfully transferred to math_expert",
"role": "tool",
"tool_call_id": "call_z9x3nnqjvnp9hxuehejkivgn"
},
{
"content": null,
"name": "math_expert",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_jd5esfvy2qsh9tdz23361pha",
"function": {
"name": "add",
"arguments": "{\"a\": 67317, \"b\": 164000}"
}
}
]
},
{
"content": "231317.0",
"role": "tool",
"tool_call_id": "call_jd5esfvy2qsh9tdz23361pha"
},
{
"content": null,
"name": "math_expert",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_owov4sx70lmdpz7zooldnue3",
"function": {
"name": "add",
"arguments": "{\"a\": 231317, \"b\": 1551000}"
}
}
]
},
{
"content": "1782317.0",
"role": "tool",
"tool_call_id": "call_owov4sx70lmdpz7zooldnue3"
}
]
}
# 大模型再次返回了調(diào)用加法工具，并提供了參數(shù)，如此循環(huán)多次
response: {
"id": "02174685631635591645265e627e21d5dcc187f9552fedf2db39e",
"choices": [
{
"finish_reason": "tool_calls",
"index": 0,
"logprobs": null,
"message": {
"content": "",
"role": "assistant",
"tool_calls": [
{
"id": "call_w2tp827rhg290ekal228ei1c",
"function": {
"arguments": " {\"a\": 1782317, \"b\": 14000}",
"name": "add"
},
"type": "function"
}
]
}
}
],
}
request: {
"tools": [
{
"type": "function",
"function": {
"name": "add",
"description": "Add two numbers.",
"parameters": {
"properties": {
"a": {
"type": "number"
},
"b": {
"type": "number"
}
},
"required": [
"a",
"b"
],
"type": "object"
}
}
},
{
"type": "function",
"function": {
"name": "multiply",
"description": "Multiply two numbers.",
"parameters": {
"properties": {
"a": {
"type": "number"
},
"b": {
"type": "number"
}
},
"required": [
"a",
"b"
],
"type": "object"
}
}
}
],
"messages": [
{
"content": "You are a math expert. Always use one tool at a time.",
"role": "system"
},
{
"content": "what's the combined headcount of the FAANG companies in 2024?",
"role": "user"
},
{
"content": null,
"name": "supervisor",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_fyew4c06vn9ide7nhk6sijws",
"function": {
"name": "transfer_to_research_expert",
"arguments": "{}"
}
}
]
},
{
"content": "Successfully transferred to research_expert",
"role": "tool",
"tool_call_id": "call_fyew4c06vn9ide7nhk6sijws"
},
{
"content": "The individual headcounts of FAANG companies in 2024 are as follows: Meta has 67,317 employees, Apple has 164,000 employees, Amazon has 1,551,000 employees, Netflix has 14,000 employees, and Alphabet has 181,269 employees. You can sum these figures to get the combined head - count. ",
"name": "research_expert",
"role": "assistant"
},
{
"content": "Transferring back to supervisor",
"name": "research_expert",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "267ff0b0-d27b-4764-a68b-71eda788379a",
"function": {
"name": "transfer_back_to_supervisor",
"arguments": "{}"
}
}
]
},
{
"content": "Successfully transferred back to supervisor",
"role": "tool",
"tool_call_id": "267ff0b0-d27b-4764-a68b-71eda788379a"
},
{
"content": null,
"name": "supervisor",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_z9x3nnqjvnp9hxuehejkivgn",
"function": {
"name": "transfer_to_math_expert",
"arguments": "{}"
}
}
]
},
{
"content": "Successfully transferred to math_expert",
"role": "tool",
"tool_call_id": "call_z9x3nnqjvnp9hxuehejkivgn"
},
{
"content": null,
"name": "math_expert",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_jd5esfvy2qsh9tdz23361pha",
"function": {
"name": "add",
"arguments": "{\"a\": 67317, \"b\": 164000}"
}
}
]
},
{
"content": "231317.0",
"role": "tool",
"tool_call_id": "call_jd5esfvy2qsh9tdz23361pha"
},
{
"content": null,
"name": "math_expert",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_owov4sx70lmdpz7zooldnue3",
"function": {
"name": "add",
"arguments": "{\"a\": 231317, \"b\": 1551000}"
}
}
]
},
{
"content": "1782317.0",
"role": "tool",
"tool_call_id": "call_owov4sx70lmdpz7zooldnue3"
},
{
"content": null,
"name": "math_expert",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_w2tp827rhg290ekal228ei1c",
"function": {
"name": "add",
"arguments": "{\"a\": 1782317, \"b\": 14000}"
}
}
]
},
{
"content": "1796317.0",
"role": "tool",
"tool_call_id": "call_w2tp827rhg290ekal228ei1c"
}
]
}
response: {
"id": "02174685631866091645265e627e21d5dcc187f9552fedf7df673",
"choices": [
{
"finish_reason": "tool_calls",
"index": 0,
"logprobs": null,
"message": {
"content": "",
"role": "assistant",
"tool_calls": [
{
"id": "call_rvoyhqsoj5sz1g0a9tz28qy9",
"function": {
"arguments": " {\"a\": 1796317, \"b\": 181269}",
"name": "add"
},
"type": "function"
}
]
}
}
],
}
# 數(shù)學(xué)計(jì)算 Agent 最后一次調(diào)用大模型
request: {
"tools": [
{
"type": "function",
"function": {
"name": "add",
"description": "Add two numbers.",
"parameters": {
"properties": {
"a": {
"type": "number"
},
"b": {
"type": "number"
}
},
"required": [
"a",
"b"
],
"type": "object"
}
}
},
{
"type": "function",
"function": {
"name": "multiply",
"description": "Multiply two numbers.",
"parameters": {
"properties": {
"a": {
"type": "number"
},
"b": {
"type": "number"
}
},
"required": [
"a",
"b"
],
"type": "object"
}
}
}
],
"messages": [
{
"content": "You are a math expert. Always use one tool at a time.",
"role": "system"
},
{
"content": "what's the combined headcount of the FAANG companies in 2024?",
"role": "user"
},
{
"content": null,
"name": "supervisor",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_fyew4c06vn9ide7nhk6sijws",
"function": {
"name": "transfer_to_research_expert",
"arguments": "{}"
}
}
]
},
{
"content": "Successfully transferred to research_expert",
"role": "tool",
"tool_call_id": "call_fyew4c06vn9ide7nhk6sijws"
},
{
"content": "The individual headcounts of FAANG companies in 2024 are as follows: Meta has 67,317 employees, Apple has 164,000 employees, Amazon has 1,551,000 employees, Netflix has 14,000 employees, and Alphabet has 181,269 employees. You can sum these figures to get the combined head - count. ",
"name": "research_expert",
"role": "assistant"
},
{
"content": "Transferring back to supervisor",
"name": "research_expert",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "267ff0b0-d27b-4764-a68b-71eda788379a",
"function": {
"name": "transfer_back_to_supervisor",
"arguments": "{}"
}
}
]
},
{
"content": "Successfully transferred back to supervisor",
"role": "tool",
"tool_call_id": "267ff0b0-d27b-4764-a68b-71eda788379a"
},
{
"content": null,
"name": "supervisor",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_z9x3nnqjvnp9hxuehejkivgn",
"function": {
"name": "transfer_to_math_expert",
"arguments": "{}"
}
}
]
},
{
"content": "Successfully transferred to math_expert",
"role": "tool",
"tool_call_id": "call_z9x3nnqjvnp9hxuehejkivgn"
},
{
"content": null,
"name": "math_expert",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_jd5esfvy2qsh9tdz23361pha",
"function": {
"name": "add",
"arguments": "{\"a\": 67317, \"b\": 164000}"
}
}
]
},
{
"content": "231317.0",
"role": "tool",
"tool_call_id": "call_jd5esfvy2qsh9tdz23361pha"
},
{
"content": null,
"name": "math_expert",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_owov4sx70lmdpz7zooldnue3",
"function": {
"name": "add",
"arguments": "{\"a\": 231317, \"b\": 1551000}"
}
}
]
},
{
"content": "1782317.0",
"role": "tool",
"tool_call_id": "call_owov4sx70lmdpz7zooldnue3"
},
{
"content": null,
"name": "math_expert",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_w2tp827rhg290ekal228ei1c",
"function": {
"name": "add",
"arguments": "{\"a\": 1782317, \"b\": 14000}"
}
}
]
},
{
"content": "1796317.0",
"role": "tool",
"tool_call_id": "call_w2tp827rhg290ekal228ei1c"
},
{
"content": null,
"name": "math_expert",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_rvoyhqsoj5sz1g0a9tz28qy9",
"function": {
"name": "add",
"arguments": "{\"a\": 1796317, \"b\": 181269}"
}
}
]
},
{
"content": "1977586.0",
"role": "tool",
"tool_call_id": "call_rvoyhqsoj5sz1g0a9tz28qy9"
}
]
}
# 大模型返回結(jié)果，計(jì)算完畢
response: {
"id": "02174685632091091645265e627e21d5dcc187f9552fedfe2a3f5",
"choices": [
{
"finish_reason": "stop",
"index": 0,
"logprobs": null,
"message": {
"content": "The combined headcount of the FAANG companies in 2024 is 1,977,586. ",
"role": "assistant"
}
}
],
}5.handoff 到 supervisor
# handoff 到 supervisor，supervisor 再次調(diào)大模型
request: {
"tools": [
{
"type": "function",
"function": {
"name": "transfer_to_research_expert",
"description": "Ask agent 'research_expert' for help",
"parameters": {
"properties": {},
"type": "object"
}
}
},
{
"type": "function",
"function": {
"name": "transfer_to_math_expert",
"description": "Ask agent 'math_expert' for help",
"parameters": {
"properties": {},
"type": "object"
}
}
}
],
"parallel_tool_calls": false,
"messages": [
{
"content": "You are a team supervisor managing a research expert and a math expert. For current events, use research_agent. For math problems, use math_agent.",
"role": "system"
},
{
"content": "what's the combined headcount of the FAANG companies in 2024?",
"role": "user"
},
{
"content": null,
"name": "supervisor",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_fyew4c06vn9ide7nhk6sijws",
"function": {
"name": "transfer_to_research_expert",
"arguments": "{}"
}
}
]
},
{
"content": "Successfully transferred to research_expert",
"role": "tool",
"tool_call_id": "call_fyew4c06vn9ide7nhk6sijws"
},
{
"content": "The individual headcounts of FAANG companies in 2024 are as follows: Meta has 67,317 employees, Apple has 164,000 employees, Amazon has 1,551,000 employees, Netflix has 14,000 employees, and Alphabet has 181,269 employees. You can sum these figures to get the combined head - count. ",
"name": "research_expert",
"role": "assistant"
},
{
"content": "Transferring back to supervisor",
"name": "research_expert",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "267ff0b0-d27b-4764-a68b-71eda788379a",
"function": {
"name": "transfer_back_to_supervisor",
"arguments": "{}"
}
}
]
},
{
"content": "Successfully transferred back to supervisor",
"role": "tool",
"tool_call_id": "267ff0b0-d27b-4764-a68b-71eda788379a"
},
{
"content": null,
"name": "supervisor",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "call_z9x3nnqjvnp9hxuehejkivgn",
"function": {
"name": "transfer_to_math_expert",
"arguments": "{}"
}
}
]
},
{
"content": "Successfully transferred to math_expert",
"role": "tool",
"tool_call_id": "call_z9x3nnqjvnp9hxuehejkivgn"
},
{
"content": "The combined headcount of the FAANG companies in 2024 is 1,977,586. ",
"name": "math_expert",
"role": "assistant"
},
{
"content": "Transferring back to supervisor",
"name": "math_expert",
"role": "assistant",
"tool_calls": [
{
"type": "function",
"id": "dff2e74c-2997-4acc-b86d-9d702f933dcb",
"function": {
"name": "transfer_back_to_supervisor",
"arguments": "{}"
}
}
]
},
{
"content": "Successfully transferred back to supervisor",
"role": "tool",
"tool_call_id": "dff2e74c-2997-4acc-b86d-9d702f933dcb"
}
]
}
# 大模型返回最終結(jié)果
response: {
"id": "02174685632249991645265e627e21d5dcc187f9552fedf38b54d",
"choices": [
{
"finish_reason": "stop",
"index": 0,
"logprobs": null,
"message": {
"content": "The combined headcount of the FAANG companies in 2024 is 1,977,586. ",
"role": "assistant"
}
}
],
}6.拿到最終結(jié)果。
責(zé)任編輯：武曉燕
來源：
編程雜技
