MetaGPT Usage & Development Guide
输入“/”快速插入内容
MetaGPT Usage & Development Guide
MetaGPT Usage & Development Guide
用户7255
用户2278
用户9941
用户1772用户722用户62232023年9月21日修改
Concept
•
Single Agent = LLM (for thinking) + Action (of using tools) + Memory
•
Multiple Agents: need an environment to interact with each other
•
Agent 101
Flowchart of one agent run cycle
画板
Usage of off-the-shelf agents
Import any role, initialize it, run it with a starting message, done!
代码块
import asyncio
from metagpt.roles.product_manager import ProductManager
from metagpt.logs import logger
async def main():
msg = "Write a PRD for a snake game"
role = ProductManager()
result = await role.run(msg)
logger.info(result.content[:100])
if __name__ == '__main__':
asyncio.run(main())
Develop your first agent
Academia and industry have proposed various definitions for the term "agent". Roughly speaking, an agent should be able to think or plan like human, possesses memories or even emotions, is equipped with a certain skill set to interact with the environment, other agents, and human. An agent, in a comprehensive examination, is a sophisticated system by itself.
However, for now, let's put completeness aside (reserving such thoroughness in Advanced Topics), tolerate a bit of simplification, and consider agent from a practical usage viewpoint. The question arises: "What are the bare essentials for an agent to be of any utility to us?" From MetaGPT's standpoint, if an agent can execute certain actions (whether powered by LLM or otherwise), it holds some degree of usefulness. Put simply, we define what actions our agent is expected to possess, equip the agent with these capabilities, and we have a basic useful agent! MetaGPT provides high flexibility to define your own action and your own agent. We will walk you through this in the rest of this section.
Agent with a single action
Suppose we want to write codes in natural language and want an agent to do this for us. Let's call this agent SimpleCoder and we need two steps to put it to work:
1.
Define a write code action
2.
Equip the agent with the action
Define Actions
In MetaGPT, class Action is the logical abstraction for an action. Users may use LLM to empower this Action by simply invoking the self._aask function, which will make LLM api call under the hood.
In our scenario, we define a SimpleWriteCode subclassed Action. Although it primarily acts as a wrapper around a prompt and the LLM call, we believe that this Action abstraction is more intuitive. In downstream and higher-level tasks, using it as a whole feels more natural than crafting a prompt and invoking the LLM separately, especially when viewed within the framework of an agent.
代码块
from metagpt.actions import Action
class SimpleWriteCode(Action):
PROMPT_TEMPLATE = """
Write a python function that can {instruction} and provide two runnnable test cases.
Quote all your code with a single ``` ```,
example:
```
# function
def add(a, b):
return a + b
# test cases
print(add(1, 2))
print(add(3, 4))
```
your code:
"""
def __init__(self, name="SimpleWriteCode", context=None, llm=None):
super().__init__(name, context, llm)
async def run(self, instruction: str):
prompt = self.PROMPT_TEMPLATE.format(instruction=instruction)
rsp = await self._aask(prompt)
code_text = SimpleWriteCode.parse_code(rsp)
return code_text