Overview

构建agent的难点在于使他们足够可靠 即时他们在原型阶段往往会表现得很好

Why do agent fails

LLM调用采取了错误的措施或者没有按照预期执行,原因:

1. LLM能力不够
2. 正确的context没有传递给LLM

更多情况是后者

The agent loop

two main step2:

1. Model call: calls the LLM with a prompt and available tools, returns either a response or a request to execute tools
2. Tool execution: execute the tools that LLM requested,returns tool results

What can you control:

Data sources

How it works

Using middleware:

• update context
• Jump to a different step in the agent lifecycle

Model context

Instructions,available tools,which model to use,output format.

• System prompt
• Messages:conversation history
• Tools
• Model
• Response format

System prompt

Successful agents draw on memories, preferences, and configuration to provide the right instructions for the current state of the conversation.

• Access message count or conversation context from state
• Access user preferences from long-term memory
• Access user ID or configuration from Runtime Context

Message

Make up the prompt that is sent to the LLM

分为state,store,Runtime context

Tools

Defining tools

name,description(docstring),argument name, argument description

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20

from langchain.tools import tool

@tool(parse_docstring=True)
def search_orders(
    user_id: str,
    status: str,
    limit: int = 10
) -> str:
    """Search for user orders by status.

    Use this when the user asks about order history or wants to check
    order status. Always filter by the provided status.

    Args:
        user_id: Unique identifier for the user
        status: Order status: 'pending', 'shipped', or 'delivered'
        limit: Maximum number of results to return
    """
    # Implementation here
    pass

Selecting tools

从分为从三个数据源来组织上下文

Model

根据三种数据源中的各种数据来选择合适的模型

Response format

defining formats

Schema definitions guide the model. Field names, types, and descriptions specify exactly what format the output should adhere to.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17

from pydantic import BaseModel, Field

class CustomerSupportTicket(BaseModel):
    """Structured ticket information extracted from customer message."""

    category: str = Field(
        description="Issue category: 'billing', 'technical', 'account', or 'product'"
    )
    priority: str = Field(
        description="Urgency level: 'low', 'medium', 'high', or 'critical'"
    )
    summary: str = Field(
        description="One-sentence summary of the customer's issue"
    )
    customer_sentiment: str = Field(
        description="Customer's emotional tone: 'frustrated', 'neutral', or 'satisfied'"
    )

selecting format

Tool context

Both read and write context

Reads

大多数的工具不只是需要LLM传入的参数,需要从其他的数据源获得,主要是state,store,Runtime context

Writes

update the memory

• Write to State to track session-specific information using Command:
• Write to Store to persist data across sessions:

Life-cycle context

控制核心agent步骤之间发生的情况:拦截数据流以实现总结,安全护栏,日志记录等横切关注点

使用middleware:

• update context:修改state和存储以持久化更改,更新对话历史或者保存insights
• Jump in the lifecycle:根据上下文移动到Agent loop中的不同步骤(例如如果满足条件就跳过工具执行,或使用修改后的上下文重复model call)

Example: Summarization

automatically condensing conversation history

LangChain offers built-in middleware for this:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14

from langchain.agents import create_agent
from langchain.agents.middleware import SummarizationMiddleware

agent = create_agent(
    model="gpt-5.5",
    tools=[...],
    middleware=[
        SummarizationMiddleware(
            model="gpt-5.4-mini",
            trigger={"tokens": 4000},
            keep=("messages", 20),
        ),
    ],
)

When the conversation exceeds the token limit, SummarizationMiddleware automatically:

1. Summarizes older messages using a separate LLM call
2. Replaces them with a summary message in State (permanently)
3. Keeps recent messages intact for context