使用按需技能构建 SQL 助手

本教程展示如何使用渐进式披露——一种智能体按需加载信息而非预先加载的上下文管理技术——来实现技能（基于提示的专业指令）。智能体通过工具调用加载技能，而非动态更改系统提示，仅发现和加载每个任务所需的技能。

**使用场景：**想象构建一个智能体来帮助在大型企业中跨不同业务垂直领域编写 SQL 查询。你的组织可能为每个垂直领域有单独的数据存储，或者有一个包含数千张表的单体数据库。无论哪种情况，预先加载所有 schema 都会淹没上下文窗口。渐进式披露通过仅在需要时加载相关 schema 来解决这个问题。这种架构还使不同的产品负责人和利益相关者能够独立贡献和维护其特定业务垂直领域的技能。

**你将构建什么：**一个具有两个技能（销售分析和库存管理）的 SQL 查询助手。智能体在其系统提示中看到轻量级的技能描述，然后仅在与用户查询相关时通过工具调用加载完整的数据库 schema 和业务逻辑。

For a complete example of a SQL agent with query execution, error correction, and validation, see our SQL Agent tutorial. This tutorial focuses on the progressive disclosure pattern which can be applied to any domain.

Progressive disclosure was popularized by Anthropic as a technique for building scalable agent skills systems. This approach uses a three-level architecture (metadata → core content → detailed resources) where agents load information only as needed. For more on this technique, see Equipping agents for the real world with Agent Skills.

工作原理

Here’s the flow when a user asks for a SQL query:

Why progressive disclosure:

• Reduces context usage - load only the 2-3 skills needed for a task, not all available skills
• Enables team autonomy - different teams can develop specialized skills independently (similar to other multi-agent architectures)
• Scales efficiently - add dozens or hundreds of skills without overwhelming context
• Simplifies conversation history - single agent with one conversation thread

What are skills: Skills, as popularized by Claude Code, are primarily prompt-based: self-contained units of specialized instructions for specific business tasks. In Claude Code, skills are exposed as directories with files on the file system, discovered through file operations. Skills guide behavior through prompts and can provide information about tool usage or include sample code for a coding agent to execute.

Skills with progressive disclosure can be viewed as a form of RAG (Retrieval-Augmented Generation), where each skill is a retrieval unit—though not necessarily backed by embeddings or keyword search, but by tools for browsing content (like file operations or, in this tutorial, direct lookup).

Trade-offs:

• Latency: Loading skills on-demand requires additional tool calls, which adds latency to the first request that needs each skill
• Workflow control: Basic implementations rely on prompting to guide skill usage - you cannot enforce hard constraints like “always try skill A before skill B” without custom logic

Implementing your own skills system

When building your own skills implementation (as we do in this tutorial), the core concept is progressive disclosure - loading information on-demand. Beyond that, you have full flexibility in implementation:

• Storage: databases, S3, in-memory data structures, or any backend
• Discovery: direct lookup (this tutorial), RAG for large skill collections, file system scanning, or API calls
• Loading logic: customize latency characteristics and add logic to search through skill content or rank relevance
• Side effects: define what happens when a skill loads, such as exposing tools associated with that skill (covered in section 8)

This flexibility lets you optimize for your specific requirements around performance, storage, and workflow control.

设置

安装

This tutorial requires the langchain package:

npm

npm install langchain

For more details, see our Installation guide.

LangSmith

Set up LangSmith to inspect what is happening inside your agent. Then set the following environment variables:

bash

export LANGSMITH_TRACING="true"
export LANGSMITH_API_KEY="..."

选择 LLM

Select a chat model from LangChain’s suite of integrations:

OpenAI

👉 Read the OpenAI chat model integration docs

npm

npm install @langchain/openai

initChatModel

import { initChatModel } from "langchain";

process.env.OPENAI_API_KEY = "your-api-key";

const model = await initChatModel("gpt-5.4");

Anthropic

👉 Read the Anthropic chat model integration docs

npm

npm install @langchain/anthropic

initChatModel

import { initChatModel } from "langchain";

process.env.ANTHROPIC_API_KEY = "your-api-key";

const model = await initChatModel("claude-sonnet-4-6");

Azure

👉 Read the Azure chat model integration docs

npm

npm install @langchain/azure

initChatModel

import { initChatModel } from "langchain";

process.env.AZURE_OPENAI_API_KEY = "your-api-key";
process.env.AZURE_OPENAI_ENDPOINT = "your-endpoint";
process.env.OPENAI_API_VERSION = "your-api-version";

const model = await initChatModel("azure_openai:gpt-5.4");

Google Gemini

👉 Read the Google GenAI chat model integration docs

npm

npm install @langchain/google-genai

initChatModel

import { initChatModel } from "langchain";

process.env.GOOGLE_API_KEY = "your-api-key";

const model = await initChatModel("google-genai:gemini-2.5-flash-lite");

Bedrock Converse

👉 Read the AWS Bedrock chat model integration docs

npm

npm install @langchain/aws

initChatModel

import { initChatModel } from "langchain";

// Follow the steps here to configure your credentials:
// https://docs.aws.amazon.com/bedrock/latest/userguide/getting-started.html

const model = await initChatModel("bedrock:gpt-5.4");

1. 定义技能

First, define the structure for skills. Each skill has a name, a brief description (shown in the system prompt), and full content (loaded on-demand):

import { z } from "zod";

// A skill that can be progressively disclosed to the agent
const SkillSchema = z.object({
  name: z.string(),  // Unique identifier for the skill
  description: z.string(),  // 1-2 sentence description to show in system prompt
  content: z.string(),  // Full skill content with detailed instructions
});

type Skill = z.infer<typeof SkillSchema>;

Now define example skills for a SQL query assistant. The skills are designed to be lightweight in description (shown to the agent upfront) but detailed in content (loaded only when needed):

View complete skill definitions

import { context } from "langchain";

const SKILLS: Skill[] = [
  {
    name: "sales_analytics",
    description:
      "Database schema and business logic for sales data analysis including customers, orders, and revenue.",
    content: context`
    # Sales Analytics Schema

    ## Tables

    ### customers
    - customer_id (PRIMARY KEY)
    - name
    - email
    - signup_date
    - status (active/inactive)
    - customer_tier (bronze/silver/gold/platinum)

    ### orders
    - order_id (PRIMARY KEY)
    - customer_id (FOREIGN KEY -> customers)
    - order_date
    - status (pending/completed/cancelled/refunded)
    - total_amount
    - sales_region (north/south/east/west)

    ### order_items
    - item_id (PRIMARY KEY)
    - order_id (FOREIGN KEY -> orders)
    - product_id
    - quantity
    - unit_price
    - discount_percent

    ## Business Logic

    **Active customers**:
    status = 'active' AND signup_date <= CURRENT_DATE - INTERVAL '90 days'

    **Revenue calculation**:
    Only count orders with status = 'completed'.
    Use total_amount from orders table, which already accounts for discounts.

    **Customer lifetime value (CLV)**:
    Sum of all completed order amounts for a customer.

    **High-value orders**:
    Orders with total_amount > 1000

    ## Example Query

    -- Get top 10 customers by revenue in the last quarter
    SELECT
        c.customer_id,
        c.name,
        c.customer_tier,
        SUM(o.total_amount) as total_revenue
    FROM customers c
    JOIN orders o ON c.customer_id = o.customer_id
    WHERE o.status = 'completed'
    AND o.order_date >= CURRENT_DATE - INTERVAL '3 months'
    GROUP BY c.customer_id, c.name, c.customer_tier
    ORDER BY total_revenue DESC
    LIMIT 10;`,
  },
  {
    name: "inventory_management",
    description:
      "Database schema and business logic for inventory tracking including products, warehouses, and stock levels.",
    content: context`
    # Inventory Management Schema

    ## Tables

    ### products
    - product_id (PRIMARY KEY)
    - product_name
    - sku
    - category
    - unit_cost
    - reorder_point (minimum stock level before reordering)
    - discontinued (boolean)

    ### warehouses
    - warehouse_id (PRIMARY KEY)
    - warehouse_name
    - location
    - capacity

    ### inventory
    - inventory_id (PRIMARY KEY)
    - product_id (FOREIGN KEY -> products)
    - warehouse_id (FOREIGN KEY -> warehouses)
    - quantity_on_hand
    - last_updated

    ### stock_movements
    - movement_id (PRIMARY KEY)
    - product_id (FOREIGN KEY -> products)
    - warehouse_id (FOREIGN KEY -> warehouses)
    - movement_type (inbound/outbound/transfer/adjustment)
    - quantity (positive for inbound, negative for outbound)
    - movement_date
    - reference_number

    ## Business Logic

    **Available stock**:
    quantity_on_hand from inventory table where quantity_on_hand > 0

    **Products needing reorder**:
    Products where total quantity_on_hand across all warehouses is less
    than or equal to the product's reorder_point

    **Active products only**:
    Exclude products where discontinued = true unless specifically analyzing discontinued items

    **Stock valuation**:
    quantity_on_hand * unit_cost for each product

    ## Example Query

    -- Find products below reorder point across all warehouses
    SELECT
        p.product_id,
        p.product_name,
        p.reorder_point,
        SUM(i.quantity_on_hand) as total_stock,
        p.unit_cost,
        (p.reorder_point - SUM(i.quantity_on_hand)) as units_to_reorder
    FROM products p
    JOIN inventory i ON p.product_id = i.product_id
    WHERE p.discontinued = false
    GROUP BY p.product_id, p.product_name, p.reorder_point, p.unit_cost
    HAVING SUM(i.quantity_on_hand) <= p.reorder_point
    ORDER BY units_to_reorder DESC;`,
  },
];

2. 创建技能加载工具

Create a tool to load full skill content on-demand:

import { tool } from "langchain";
import { z } from "zod";

const loadSkill = tool(  
  async ({ skillName }) => {
    // Find and return the requested skill
    const skill = SKILLS.find((s) => s.name === skillName);
    if (skill) {
      return `Loaded skill: ${skillName}\n\n${skill.content}`;
    }

    // Skill not found
    const available = SKILLS.map((s) => s.name).join(", ");
    return `Skill '${skillName}' not found. Available skills: ${available}`;
  },
  {
    name: "load_skill",
    description: `Load the full content of a skill into the agent's context.

Use this when you need detailed information about how to handle a specific
type of request. This will provide you with comprehensive instructions,
policies, and guidelines for the skill area.`,
    schema: z.object({
      skillName: z.string().describe("The name of the skill to load"),
    }),
  }
);

The load_skill tool returns the full skill content as a string, which becomes part of the conversation as a ToolMessage. For more details on creating and using tools, see the Tools guide.

3. 构建技能中间件

Create custom middleware that injects skill descriptions into the system prompt. This middleware makes skills discoverable without loading their full content upfront.

This guide demonstrates creating custom middleware. For a comprehensive guide on middleware concepts and patterns, see the custom middleware documentation.

import { createMiddleware } from "langchain";

// Build skills prompt from the SKILLS list
const skillsPrompt = SKILLS.map(
  (skill) => `- **${skill.name}**: ${skill.description}`
).join("\n");

const skillMiddleware = createMiddleware({
  name: "skillMiddleware",
  tools: [loadSkill],
  wrapModelCall: async (request, handler) => {
    // Build the skills addendum
    const skillsAddendum =
      `\n\n## Available Skills\n\n${skillsPrompt}\n\n` +
      "Use the load_skill tool when you need detailed information " +
      "about handling a specific type of request.";

    // Append to system prompt
    const newSystemPrompt = request.systemPrompt + skillsAddendum;

    return handler({
      ...request,
      systemPrompt: newSystemPrompt,
    });
  },
});

The middleware appends skill descriptions to the system prompt, making the agent aware of available skills without loading their full content. The load_skill tool is registered as a class variable, making it available to the agent.

Production consideration: This tutorial loads the skill list in __init__ for simplicity. In a production system, you may want to load skills in the before_agent hook instead, allowing them to be refreshed periodically to reflect up-to-date changes (e.g., when new skills are added or existing ones are modified). See the before_agent hook documentation for details.

4. 创建带技能支持的智能体

Now create the agent with the skill middleware and a checkpointer for state persistence:

import { createAgent } from "langchain";
import { MemorySaver } from "@langchain/langgraph";

// Create the agent with skill support
const agent = createAgent({
  model,
  systemPrompt:
    "You are a SQL query assistant that helps users " +
    "write queries against business databases.",
  middleware: [skillMiddleware],
  checkpointer: new MemorySaver(),
});

The agent now has access to skill descriptions in its system prompt and can call load_skill to retrieve full skill content when needed. The checkpointer maintains conversation history across turns.

5. 测试渐进式披露

Test the agent with a question that requires skill-specific knowledge:

import { v4 as uuidv4 } from "uuid";

// Configuration for this conversation thread
const threadId = uuidv4();
const config = { configurable: { thread_id: threadId } };

// Ask for a SQL query
const result = await agent.invoke(  
  {
    messages: [
      {
        role: "user",
        content:
          "Write a SQL query to find all customers " +
          "who made orders over $1000 in the last month",
      },
    ],
  },
  config
);

// Print the conversation
for (const message of result.messages) {
  console.log(`${message._getType()}: ${message.content}`);
}

Expected output:

================================ Human Message =================================

Write a SQL query to find all customers who made orders over $1000 in the last month
================================== Ai Message ==================================
Tool Calls:
  load_skill (call_abc123)
 Call ID: call_abc123
  Args:
    skill_name: sales_analytics
================================= Tool Message =================================
Name: load_skill

Loaded skill: sales_analytics

# Sales Analytics Schema

## Tables

### customers
- customer_id (PRIMARY KEY)
- name
- email
- signup_date
- status (active/inactive)
- customer_tier (bronze/silver/gold/platinum)

### orders
- order_id (PRIMARY KEY)
- customer_id (FOREIGN KEY -> customers)
- order_date
- status (pending/completed/cancelled/refunded)
- total_amount
- sales_region (north/south/east/west)

[... rest of schema ...]

## Business Logic

**High-value orders**: Orders with `total_amount > 1000`
**Revenue calculation**: Only count orders with `status = 'completed'`

================================== Ai Message ==================================

Here's a SQL query to find all customers who made orders over $1000 in the last month:

\`\`\`sql
SELECT DISTINCT
    c.customer_id,
    c.name,
    c.email,
    c.customer_tier
FROM customers c
JOIN orders o ON c.customer_id = o.customer_id
WHERE o.total_amount > 1000
  AND o.status = 'completed'
  AND o.order_date >= CURRENT_DATE - INTERVAL '1 month'
ORDER BY c.customer_id;
\`\`\`

This query:
- Joins customers with their orders
- Filters for high-value orders (>$1000) using the total_amount field
- Only includes completed orders (as per the business logic)
- Restricts to orders from the last month
- Returns distinct customers to avoid duplicates if they made multiple qualifying orders

The agent saw the lightweight skill description in its system prompt, recognized the question required sales database knowledge, called load_skill("sales_analytics") to get the full schema and business logic, and then used that information to write a correct query following the database conventions.

6. 进阶：使用自定义状态添加约束

Optional: Track loaded skills and enforce tool constraints

You can add constraints to enforce that certain tools are only available after specific skills have been loaded. This requires tracking which skills have been loaded in custom agent state.

定义自定义状态

First, extend the agent state to track loaded skills:

import { StateSchema } from "@langchain/langgraph";
import { z } from "zod";

const CustomState = new StateSchema({
  skillsLoaded: z.array(z.string()).optional(),  // Track which skills have been loaded
});

更新 load_skill 以修改状态

Modify the load_skill tool to update state when a skill is loaded:

import { tool, ToolMessage, type ToolRuntime } from "langchain";
import { Command } from "@langchain/langgraph";
import { z } from "zod";

const loadSkill = tool(  
  async ({ skillName }, runtime: ToolRuntime<typeof CustomState.State>) => {
    // Find and return the requested skill
    const skill = SKILLS.find((s) => s.name === skillName);

    if (skill) {
      const skillContent = `Loaded skill: ${skillName}\n\n${skill.content}`;

      // Update state to track loaded skill
      return new Command({
        update: {
          messages: [  
            new ToolMessage({
              content: skillContent,
              tool_call_id: runtime.toolCallId,
            }),
          ],
          skillsLoaded: [skillName],
        },
      });
    }

    // Skill not found
    const available = SKILLS.map((s) => s.name).join(", ");
    return new Command({
      update: {
        messages: [
          new ToolMessage({
            content: `Skill '${skillName}' not found. Available skills: ${available}`,
            tool_call_id: runtime.toolCallId,
          }),
        ],
      },
    });
  },
  {
    name: "load_skill",
    description: `Load the full content of a skill into the agent's context.`,
    schema: z.object({
      skillName: z.string().describe("The name of the skill to load"),
    }),
  }
);

创建受约束的工具

Create a tool that’s only usable after a specific skill has been loaded:

const writeSqlQuery = tool(  
  async ({ query, vertical }, runtime: ToolRuntime<typeof CustomState.State>) => {
    // Check if the required skill has been loaded
    const skillsLoaded = runtime.state.skillsLoaded ?? [];

    if (!skillsLoaded.includes(vertical)) {
      return (  
        `Error: You must load the '${vertical}' skill first ` +
        `to understand the database schema before writing queries. ` +
        `Use load_skill('${vertical}') to load the schema.`
      );
    }

    // Validate and format the query
    return (
      `SQL Query for ${vertical}:\n\n` +
      `\`\`\`sql\n${query}\n\`\`\`\n\n` +
      `✓ Query validated against ${vertical} schema\n` +
      `Ready to execute against the database.`
    );
  },
  {
    name: "write_sql_query",
    description: `Write and validate a SQL query for a specific business vertical.

This tool helps format and validate SQL queries. You must load the
appropriate skill first to understand the database schema.`,
    schema: z.object({
      query: z.string().describe("The SQL query to write"),
      vertical: z.string().describe("The business vertical (sales_analytics or inventory_management)"),
    }),
  }
);

更新中间件和智能体

Update the middleware to use the custom state schema:

const skillMiddleware = createMiddleware({
  name: "skillMiddleware",
  stateSchema: CustomState,
  tools: [loadSkill, writeSqlQuery],
  // ... rest of the middleware implementation stays the same
});

Create the agent with the middleware that registers the constrained tool:

const agent = createAgent({
  model,
  systemPrompt:
    "You are a SQL query assistant that helps users " +
    "write queries against business databases.",
  middleware: [skillMiddleware],
  checkpointer: new MemorySaver(),
});

Now if the agent tries to use write_sql_query before loading the required skill, it will receive an error message prompting it to load the appropriate skill (e.g., sales_analytics or inventory_management) first. This ensures the agent has the necessary schema knowledge before attempting to validate queries.

完整示例

View complete runnable script

Here’s a complete, runnable implementation combining all the pieces from this tutorial:

import {
  tool,
  createAgent,
  createMiddleware,
  ToolMessage,
  context,
  type ToolRuntime,
} from "langchain";
import { MemorySaver, Command } from "@langchain/langgraph";
import { ChatOpenAI } from "@langchain/openai";
import { v4 as uuidv4 } from "uuid";
import { z } from "zod";

// A skill that can be progressively disclosed to the agent
const SkillSchema = z.object({
  name: z.string(), // Unique identifier for the skill
  description: z.string(), // 1-2 sentence description to show in system prompt
  content: z.string(), // Full skill content with detailed instructions
});

type Skill = z.infer<typeof SkillSchema>;

const SKILLS: Skill[] = [
  {
    name: "sales_analytics",
    description:
      "Database schema and business logic for sales data analysis including customers, orders, and revenue.",
    content: context`
    # Sales Analytics Schema

    ## Tables

    ### customers
    - customer_id (PRIMARY KEY)
    - name
    - email
    - signup_date
    - status (active/inactive)
    - customer_tier (bronze/silver/gold/platinum)

    ### orders
    - order_id (PRIMARY KEY)
    - customer_id (FOREIGN KEY -> customers)
    - order_date
    - status (pending/completed/cancelled/refunded)
    - total_amount
    - sales_region (north/south/east/west)

    ### order_items
    - item_id (PRIMARY KEY)
    - order_id (FOREIGN KEY -> orders)
    - product_id
    - quantity
    - unit_price
    - discount_percent

    ## Business Logic

    **Active customers**: status = 'active' AND signup_date <= CURRENT_DATE - INTERVAL '90 days'

    **Revenue calculation**:
    Only count orders with status = 'completed'. Use total_amount from orders table,
    which already accounts for discounts.

    **Customer lifetime value (CLV)**:
    Sum of all completed order amounts for a customer.

    **High-value orders**:
    Orders with total_amount > 1000

    ## Example Query
    -- Get top 10 customers by revenue in the last quarter
    SELECT
        c.customer_id,
        c.name,
        c.customer_tier,
        SUM(o.total_amount) as total_revenue
    FROM customers c
    JOIN orders o ON c.customer_id = o.customer_id
    WHERE o.status = 'completed'
    AND o.order_date >= CURRENT_DATE - INTERVAL '3 months'
    GROUP BY c.customer_id, c.name, c.customer_tier
    ORDER BY total_revenue DESC
    LIMIT 10;`,
  },
  {
    name: "inventory_management",
    description:
      "Database schema and business logic for inventory tracking including products, warehouses, and stock levels.",
    content: context`
    # Inventory Management Schema

    ## Tables

    ### products
    - product_id (PRIMARY KEY)
    - product_name
    - sku
    - category
    - unit_cost
    - reorder_point (minimum stock level before reordering)
    - discontinued (boolean)

    ### warehouses
    - warehouse_id (PRIMARY KEY)
    - warehouse_name
    - location
    - capacity

    ### inventory
    - inventory_id (PRIMARY KEY)
    - product_id (FOREIGN KEY -> products)
    - warehouse_id (FOREIGN KEY -> warehouses)
    - quantity_on_hand
    - last_updated

    ### stock_movements
    - movement_id (PRIMARY KEY)
    - product_id (FOREIGN KEY -> products)
    - warehouse_id (FOREIGN KEY -> warehouses)
    - movement_type (inbound/outbound/transfer/adjustment)
    - quantity (positive for inbound, negative for outbound)
    - movement_date
    - reference_number

    ## Business Logic

    **Available stock**:
    quantity_on_hand from inventory table where quantity_on_hand > 0

    **Products needing reorder**:
    Products where total quantity_on_hand across all warehouses is
    less than or equal to the product's reorder_point

    **Active products only**:
    Exclude products where discontinued = true unless specifically
    analyzing discontinued items

    **Stock valuation**:
    quantity_on_hand * unit_cost for each product

    ## Example Query

    -- Find products below reorder point across all warehouses
    SELECT
        p.product_id,
        p.product_name,
        p.reorder_point,
        SUM(i.quantity_on_hand) as total_stock,
        p.unit_cost,
        (p.reorder_point - SUM(i.quantity_on_hand)) as units_to_reorder
    FROM products p
    JOIN inventory i ON p.product_id = i.product_id
    WHERE p.discontinued = false
    GROUP BY p.product_id, p.product_name, p.reorder_point, p.unit_cost
    HAVING SUM(i.quantity_on_hand) <= p.reorder_point
    ORDER BY units_to_reorder DESC;`,
  },
];

// const loadSkill = tool(
//   async ({ skillName }) => {
//     // Find and return the requested skill
//     const skill = SKILLS.find((s) => s.name === skillName);
//     if (skill) {
//       return `Loaded skill: ${skillName}\n\n${skill.content}`;
//     }

//     // Skill not found
//     const available = SKILLS.map((s) => s.name).join(", ");
//     return `Skill '${skillName}' not found. Available skills: ${available}`;
//   },
//   {
//     name: "load_skill",
//     description: `Load the full content of a skill into the agent's context.

// Use this when you need detailed information about how to handle a specific
// type of request. This will provide you with comprehensive instructions,
// policies, and guidelines for the skill area.`,
//     schema: z.object({
//       skillName: z.string().describe("The name of the skill to load"),
//     }),
//   }
// );

// Build skills prompt from the SKILLS list
const skillsPrompt = SKILLS.map(
  (skill) => `- **${skill.name}**: ${skill.description}`
).join("\n");

const skillMiddleware = createMiddleware({
  name: "skillMiddleware",
  tools: [loadSkill],
  wrapModelCall: async (request, handler) => {
    // Build the skills addendum
    const skillsAddendum =
      `\n\n## Available Skills\n\n${skillsPrompt}\n\n` +
      "Use the load_skill tool when you need detailed information " +
      "about handling a specific type of request.";

    // Append to system prompt
    const newSystemPrompt = request.systemPrompt + skillsAddendum;

    return handler({
      ...request,
      systemPrompt: newSystemPrompt,
    });
  },
});

const model = new ChatOpenAI({
  model: "gpt-5.4-mini",
  temperature: 0,
});

// Create the agent with skill support
const agent = createAgent({
  model,
  systemPrompt:
    "You are a SQL query assistant that helps users " +
    "write queries against business databases.",
  middleware: [skillMiddleware],
  checkpointer: new MemorySaver(),
});

// Configuration for this conversation thread
const threadId = uuidv4();
const config = { configurable: { thread_id: threadId } };

// Ask for a SQL query
const result = await agent.invoke(
  {
    messages: [
      {
        role: "user",
        content:
          "Write a SQL query to find all customers " +
          "who made orders over $1000 in the last month",
      },
    ],
  },
  config
);

// Print the conversation
for (const message of result.messages) {
  console.log(`${message.type}: ${message.content}`);
}

This complete example includes:

• Skill definitions with full database schemas
• The load_skill tool for on-demand loading
• SkillMiddleware that injects skill descriptions into the system prompt
• Agent creation with middleware and checkpointer
• Example usage showing how the agent loads skills and writes SQL queries

To run this, you’ll need to:

1. Install required packages: pip install langchain langchain-openai langgraph
2. Set your API key (e.g., export OPENAI_API_KEY=...)
3. Replace the model initialization with your preferred LLM provider

实现变体

View implementation options and trade-offs

This tutorial implemented skills as in-memory Python dictionaries loaded through tool calls. However, there are several ways to implement progressive disclosure with skills:

Storage backends:

• In-memory (this tutorial): Skills defined as Python data structures, fast access, no I/O overhead
• File system (Claude Code approach): Skills as directories with files, discovered via file operations like read_file
• Remote storage: Skills in S3, databases, Notion, or APIs, fetched on-demand

Skill discovery (how the agent learns which skills exist):

• System prompt listing: Skill descriptions in system prompt (used in this tutorial)
• File-based: Discover skills by scanning directories (Claude Code approach)
• Registry-based: Query a skill registry service or API for available skills
• Dynamic lookup: List available skills via a tool call

Progressive disclosure strategies (how skill content is loaded):

• Single load: Load entire skill content in one tool call (used in this tutorial)
• Paginated: Load skill content in multiple pages/chunks for large skills
• Search-based: Search within a specific skill’s content for relevant sections (e.g., using grep/read operations on skill files)
• Hierarchical: Load skill overview first, then drill into specific subsections

Size considerations (uncalibrated mental model - optimize for your system):

• Small skills (< 1K tokens / ~750 words): Can be included directly in system prompt and cached with prompt caching for cost savings and faster responses
• Medium skills (1-10K tokens / ~750-7.5K words): Benefit from on-demand loading to avoid context overhead (this tutorial)
• Large skills (> 10K tokens / ~7.5K words, or > 5-10% of context window): Should use progressive disclosure techniques like pagination, search-based loading, or hierarchical exploration to avoid consuming excessive context

The choice depends on your requirements: in-memory is fastest but requires redeployment for skill updates, while file-based or remote storage enables dynamic skill management without code changes.

渐进式披露与上下文工程

Combining with few-shot prompting and other techniques

Progressive disclosure is fundamentally a context engineering technique - you’re managing what information is available to the agent and when. This tutorial focused on loading database schemas, but the same principles apply to other types of context.

与少样本提示结合

For the SQL query use case, you could extend progressive disclosure to dynamically load few-shot examples that match the user’s query:

Example approach:

1. User asks: “Find customers who haven’t ordered in 6 months”
2. Agent loads sales_analytics schema (as shown in this tutorial)
3. Agent also loads 2-3 relevant example queries (via semantic search or tag-based lookup):
   • Query for finding inactive customers
   • Query with date-based filtering
   • Query joining customers and orders tables
4. Agent writes query using both schema knowledge AND example patterns

This combination of progressive disclosure (loading schemas on-demand) and dynamic few-shot prompting (loading relevant examples) creates a powerful context engineering pattern that scales to large knowledge bases while providing high-quality, grounded outputs.

后续步骤

• Learn about middleware for more dynamic agent behaviors
• Explore context engineering techniques for managing agent context
• Explore the handoffs pattern for sequential workflows
• Read the subagents pattern for parallel task routing
• See multi-agent patterns for other approaches to specialized agents
• Use LangSmith to debug and monitor skill loading

---

将这些文档连接到 Claude、VSCode 等，通过 MCP 获取实时答案。

在 GitHub 上编辑此页面或提交问题。