Quick Start

Here is a simple user guide.

Step0: Configuration

start es
set env APIKEY / host
python -m model_service -port 8000 -config simple/path

Step1: Own an agent

Assume you have a runnable agent code. Here, we use a basic LLM combined with a simple react framework including three tools(code, web_search, terminate) as an example.

class MxcAgent(BaseModel):
    llm: BaseLLM | None = Field(default=None)
    max_steps: int = Field(default=10)
    tools: List[BaseTool] = Field(default_factory=list)

    def think(self, query: str, **kwargs) -> bool:
        ...

    def act(self, **kwargs):
        ...

    def run(self, query: str, **kwargs) -> List[Message]:
        messages: List[Message] = []

        for i in range(self.max_steps):
            should_act: bool = self.think(query, messages=messages, **kwargs)
            if should_act:
                self.act(messages=messages, **kwargs)
            else:
                break
        return messages

    
query = "Analyze Xiaomi Corporation
agent = MxcAgent(llm=OpenAICompatibleBaseLLM(model_name="qwen3-32b", temperature=0.0001),
                 max_steps=10,
                 tools=[CodeTool(), DashscopeSearchTool(), TerminateTool()])
messages = agent.run(query=query)
answer = messages[-1].content
print(answer)

Step2: Implement AgentWrapper

In order to utilize the context generator and summarizer capabilities of beyondagent, please inherit from MxcAgent and BaseAgentWrapperMixin to implement the AgentWrapper.

Here, you need to customize two parts:

how to integrate the content message(insight) generated by the self.context_generator into the context.
implement the execute function to output the trajectory.

Below is a simple example of integrating trajectory-level insight into the context.

from beyondagent.core.module.agent_wrapper.base_agent_wrapper import BaseAgentWrapperMixin

class MxcAgentWrapper(MxcAgent, BaseAgentWrapperMixin):
    def execute(self, query: str, **kwargs) -> Trajectory:
        trajectory = Trajectory(steps=messages, query=query)
        context_msg = self.context_generator.execute(trajectory=trajectory)
        new_query = f"""
previous insight:
{context_msg.content}
Please consider the helpful parts from these in answering the question, to make the response more comprehensive and substantial.

user query:
{query}
        """.strip()
        
        messages = self.run(new_query, **kwargs)
        return Trajectory(query=query, steps=messages, answer=messages[-1].content, done=True)

Step3: Run AgentRunner with insight

Once you have completed the implementation of the AgentWrapper class, you will be able to utilize the capabilities of beyondagent.

Here is an example using SimpleAgentRunner. We first executed two historical tasks, then summarized the experience and made it persistent. Finally, we utilized the historical experience in a new task.

insights demo

from beyondagent.core.module.runner.simple_agent_runner import SimpleAgentRunner



mxc_agent_wrapper = MxcAgentWrapper(llm=OpenAICompatibleBaseLLM(model_name="qwen3-32b", temperature=0.0001),
                                    max_steps=10,
                                    tools=[CodeTool(), DashscopeSearchTool(), TerminateTool()])
agent_runner = SimpleAgentRunner(agent_wrapper=mxc_agent_wrapper, summarizer="default", context_generator="default")

# historical tasks
agent_runner.rollout_trajectory(query="Analyze the company Tesla.")
agent_runner.rollout_trajectory(query="Analyze the company Apple.")

# summary insights and store them
agent_runner.summary_and_store()

# run agent with historical insights
trajectory = agent_runner.rollout_trajectory(query="Analyze the company Xiaomi Corporation.")

Step4: Evaluation(Optional)

If we have a reward function that allows us to compare the performance before and after adding context, we can try this part.

Use run_agent to obtain the answer from the original agent (answer1), and use run_agent_wrapper to get the answer with added insights and experience (answer2).

Here, the reward function is used to compare and score the two answers. The reward.reward_value indicates the win rate of answer2.

# task
query = "Analyze Xiaomi Corporation."

# run agent
agent = MxcAgent(llm=OpenAICompatibleBaseLLM(model_name="qwen3-32b", temperature=0.0001),
                 max_steps=10,
                 tools=[CodeTool(), DashscopeSearchTool(), TerminateTool()])
messages = agent.run(query=query)
answer1 = messages[-1].content

# agent runner: Assume we already have some historical experience.
mxc_agent_wrapper = MxcAgentWrapper(llm=OpenAICompatibleBaseLLM(model_name="qwen3-32b", temperature=0.0001),
                                    max_steps=10,
                                    tools=[CodeTool(), DashscopeSearchTool(), TerminateTool()])
agent_runner = SimpleAgentRunner(agent_wrapper=mxc_agent_wrapper, summarizer="default", context_generator="default")
trajectory = agent_runner.rollout_trajectory(query=query)
answer2 = trajectory.answer

# pair-wise LLM evaluation
from beyondagent.core.module.reward_fn.simple_reward_fn import SimpleRewardFn
reward_fn = SimpleRewardFn(llm=OpenAICompatibleBaseLLM(model_name="qwen3-32b", temperature=0.0001))
reward = reward_fn.execute(query=query, answer1=answer1, answer2=answer2, eval_times=5)
print(f"final reward={reward.reward_value}")

5.4 KiB Raw Blame History