agentic-development（Agent 发展）

Agentic Development Skill

For building autonomous AI agents that perform multi-step tasks with tools.

Sources: Claude Agent SDK | Anthropic Claude Code Best Practices | Pydantic AI | Google Gemini Agent Development | OpenAI Building Agents

Framework Selection by Language

Python: Pydantic AI (Default)

from pydantic_ai import Agent
from pydantic import BaseModel

class SearchResult(BaseModel):
    title: str
    url: str
    summary: str

agent = Agent(
    'claude-sonnet-4-20250514',
    result_type=list[SearchResult],
    system_prompt='You are a research assistant.',
)

# Type-safe result
result = await agent.run('Find articles about AI agents')
for item in result.data:
    print(f"{item.title}: {item.url}")

Node.js / Next.js: Claude Agent SDK (Default)

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic();

// Define tools
const tools: Anthropic.Tool[] = [
  {
    name: "web_search",
    description: "Search the web for information",
    input_schema: {
      type: "object",
      properties: {
        query: { type: "string", description: "Search query" },
      },
      required: ["query"],
    },
  },
];

// Agentic loop
async function runAgent(prompt: string) {
  const messages: Anthropic.MessageParam[] = [
    { role: "user", content: prompt },
  ];

  while (true) {
    const response = await client.messages.create({
      model: "claude-sonnet-4-20250514",
      max_tokens: 4096,
      tools,
      messages,
    });

    // Check for tool use
    if (response.stop_reason === "tool_use") {
      const toolUse = response.content.find((b) => b.type === "tool_use");
      if (toolUse) {
        const result = await executeTool(toolUse.name, toolUse.input);
        messages.push({ role: "assistant", content: response.content });
        messages.push({
          role: "user",
          content: [{ type: "tool_result", tool_use_id: toolUse.id, content: result }],
        });
        continue;
      }
    }

    // Done - return final response
    return response.content.find((b) => b.type === "text")?.text;
  }
}

Core Principle

Plan first, act incrementally, verify always.

Agents that research and plan before executing consistently outperform those that jump straight to action. Break complex tasks into verifiable steps, use tools judiciously, and maintain clear state throughout execution.

Agent Architecture

Three Components (OpenAI)

┌─────────────────────────────────────────────────┐
│                    AGENT                        │
├─────────────────────────────────────────────────┤
│  Model (Brain)      │ LLM for reasoning &       │
│                     │ decision-making           │
├─────────────────────┼───────────────────────────┤
│  Tools (Arms/Legs)  │ APIs, functions, external │
│                     │ systems for action        │
├─────────────────────┼───────────────────────────┤
│  Instructions       │ System prompts defining   │
│  (Rules)            │ behavior & boundaries     │
└─────────────────────┴───────────────────────────┘

Project Structure

project/
├── src/
│   ├── agents/
│   │   ├── orchestrator.ts    # Main agent coordinator
│   │   ├── specialized/       # Task-specific agents
│   │   │   ├── researcher.ts
│   │   │   ├── coder.ts
│   │   │   └── reviewer.ts
│   │   └── base.ts            # Shared agent interface
│   ├── tools/
│   │   ├── definitions/       # Tool schemas
│   │   ├── implementations/   # Tool logic
│   │   └── registry.ts        # Tool discovery
│   ├── prompts/
│   │   ├── system/            # Agent instructions
│   │   └── templates/         # Task templates
│   └── memory/
│       ├── conversation.ts    # Short-term context
│       └── persistent.ts      # Long-term storage
├── tests/
│   ├── agents/                # Agent behavior tests
│   ├── tools/                 # Tool unit tests
│   └── evals/                 # End-to-end evaluations
└── skills/                    # Agent skills (Anthropic pattern)
    ├── skill-name/
    │   ├── instructions.md
    │   ├── scripts/
    │   └── resources/

Workflow Pattern: Explore-Plan-Execute-Verify

1. Explore Phase

// Gather context before acting
async function explore(task: Task): Promise<Context> {
  const relevantFiles = await agent.searchCodebase(task.query);
  const existingPatterns = await agent.analyzePatterns(relevantFiles);
  const dependencies = await agent.identifyDependencies(task);

  return { relevantFiles, existingPatterns, dependencies };
}

2. Plan Phase (Critical)

// Plan explicitly before execution
async function plan(task: Task, context: Context): Promise<Plan> {
  const prompt = `
    Task: ${task.description}
    Context: ${JSON.stringify(context)}

    Create a step-by-step plan. For each step:
    1. What action to take
    2. What tools to use
    3. How to verify success
    4. What could go wrong

    Output JSON with steps array.
  `;

  return await llmCall({ prompt, schema: PlanSchema });
}

3. Execute Phase

// Execute with verification at each step
async function execute(plan: Plan): Promise<Result[]> {
  const results: Result[] = [];

  for (const step of plan.steps) {
    // Execute single step
    const result = await executeStep(step);

    // Verify before continuing
    if (!await verify(step, result)) {
      // Self-correct or escalate
      const corrected = await selfCorrect(step, result);
      if (!corrected.success) {
        return handleFailure(step, results);
      }
    }

    results.push(result);
  }

  return results;
}

4. Verify Phase

// Independent verification prevents overfitting
async function verify(step: Step, result: Result): Promise<boolean> {
  // Run tests if available
  if (step.testCommand) {
    const testResult = await runCommand(step.testCommand);
    if (!testResult.success) return false;
  }

  // Use LLM to verify against criteria
  const verification = await llmCall({
    prompt: `
      Step: ${step.description}
      Expected: ${step.successCriteria}
      Actual: ${JSON.stringify(result)}

      Does the result satisfy the success criteria?
      Respond with { "passes": boolean, "reasoning": string }
    `,
    schema: VerificationSchema
  });

  return verification.passes;
}

Tool Design

Tool Definition Pattern

// tools/definitions/file-operations.ts
import { z } from 'zod';

export const ReadFileTool = {
  name: 'read_file',
  description: 'Read contents of a file. Use before modifying any file.',
  parameters: z.object({
    path: z.string().describe('Absolute path to the file'),
    startLine: z.number().optional().describe('Start line (1-indexed)'),
    endLine: z.number().optional().describe('End line (1-indexed)'),
  }),
  // Risk level for guardrails (OpenAI pattern)
  riskLevel: 'low' as const,
};

export const WriteFileTool = {
  name: 'write_file',
  description: 'Write content to a file. Always read first to understand context.',
  parameters: z.object({
    path: z.string().describe('Absolute path to the file'),
    content: z.string().describe('Complete file content'),
  }),
  riskLevel: 'medium' as const,
  // Require confirmation for high-risk operations
  requiresConfirmation: true,
};

Tool Implementation

// tools/implementations/file-operations.ts
export async function readFile(
  params: z.infer<typeof ReadFileTool.parameters>
): Promise<ToolResult> {
  try {
    const content = await fs.readFile(params.path, 'utf-8');
    const lines = content.split('\n');

    const start = (params.startLine ?? 1) - 1;
    const end = params.endLine ?? lines.length;

    return {
      success: true,
      data: lines.slice(start, end).join('\n'),
      metadata: { totalLines: lines.length }
    };
  } catch (error) {
    return {
      success: false,
      error: `Failed to read file: ${error.message}`
    };
  }
}

Prefer Built-in Tools (OpenAI)

// Use platform-provided tools when available
const agent = createAgent({
  tools: [
    // Built-in tools (handled by platform)
    { type: 'web_search' },
    { type: 'code_interpreter' },

    // Custom tools only when needed
    { type: 'function', function: customDatabaseTool },
  ],
});

Multi-Agent Patterns

Single Agent (Default)

Use one agent for most tasks. Multiple agents add complexity.

Agent-as-Tool Pattern (OpenAI)

// Expose specialized agents as callable tools
const researchAgent = createAgent({
  name: 'researcher',
  instructions: 'You research topics and return structured findings.',
  tools: [webSearchTool, documentReadTool],
});

const mainAgent = createAgent({
  tools: [
    {
      type: 'function',
      function: {
        name: 'research_topic',
        description: 'Delegate research to specialized agent',
        parameters: ResearchQuerySchema,
        handler: async (query) => researchAgent.run(query),
      },
    },
  ],
});

Handoff Pattern (OpenAI)

// One-way transfer between agents
const customerServiceAgent = createAgent({
  tools: [
    // Handoff to specialist when needed
    {
      name: 'transfer_to_billing',
      description: 'Transfer to billing specialist for payment issues',
      handler: async (context) => {
        return { handoff: 'billing_agent', context };
      },
    },
  ],
});

When to Use Multiple Agents

• Separate task domains with non-overlapping tools
• Different authorization levels needed
• Complex workflows with clear handoff points
• Parallel execution of independent subtasks

Memory & State

Conversation Memory

// memory/conversation.ts
interface ConversationMemory {
  messages: Message[];
  maxTokens: number;

  add(message: Message): void;
  getContext(): Message[];
  summarize(): Promise<string>;
}

// Maintain state across tool calls (Gemini pattern)
interface AgentState {
  thoughtSignature?: string;  // Encrypted reasoning state
  conversationId: string;     // For shared memory
  currentPlan?: Plan;
  completedSteps: Step[];
}

Persistent Memory

// memory/persistent.ts
interface PersistentMemory {
  // Store learnings across sessions
  store(key: string, value: any): Promise<void>;
  retrieve(key: string): Promise<any>;

  // Semantic search over past interactions
  search(query: string, limit: number): Promise<Memory[]>;
}

Guardrails & Safety

Multi-Layer Protection (OpenAI)

// guards/index.ts
interface GuardrailConfig {
  // Input validation
  inputClassifier: (input: string) => Promise<SafetyResult>;

  // Output validation
  outputValidator: (output: string) => Promise<SafetyResult>;

  // Tool risk assessment
  toolRiskLevels: Record<string, 'low' | 'medium' | 'high'>;

  // Actions requiring human approval
  humanInTheLoop: string[];
}

async function executeWithGuardrails(
  agent: Agent,
  input: string,
  config: GuardrailConfig
): Promise<Result> {
  // 1. Check input safety
  const inputCheck = await config.inputClassifier(input);
  if (!inputCheck.safe) {
    return { blocked: true, reason: inputCheck.reason };
  }

  // 2. Execute with tool monitoring
  const result = await agent.run(input, {
    beforeTool: async (tool, params) => {
      const risk = config.toolRiskLevels[tool.name];
      if (risk === 'high' || config.humanInTheLoop.includes(tool.name)) {
        return await requestHumanApproval(tool, params);
      }
      return { approved: true };
    },
  });

  // 3. Validate output
  const outputCheck = await config.outputValidator(result.output);
  if (!outputCheck.safe) {
    return { blocked: true, reason: outputCheck.reason };
  }

  return result;
}

Scope Enforcement (OpenAI)

// Agent must stay within defined scope
const agentInstructions = `
You are a customer service agent for Acme Corp.

SCOPE BOUNDARIES (non-negotiable):
- Only answer questions about Acme products and services
- Never provide legal, medical, or financial advice
- Never access or modify data outside your authorized scope
- If a request is out of scope, politely decline and explain why

If you cannot complete a task within scope, notify the user
and request explicit approval before proceeding.
`;

Model Selection

Match Model to Task

Gemini-Specific

// Use thinking_level for reasoning depth
const response = await gemini.generate({
  model: 'gemini-3',
  thinking_level: 'high',  // For complex planning
  temperature: 1.0,        // Optimized for reasoning engine
});

// Preserve thought state across tool calls
const nextResponse = await gemini.generate({
  thoughtSignature: response.thoughtSignature,  // Required for function calling
  // ... rest of params
});

Claude-Specific (Thinking Modes)

// Trigger extended thinking with keywords
const thinkingLevels = {
  'think': 'standard analysis',
  'think hard': 'deeper reasoning',
  'think harder': 'extensive analysis',
  'ultrathink': 'maximum reasoning budget',
};

const prompt = `
Think hard about this problem before proposing a solution.

Task: ${task.description}
`;

Testing Agents

Unit Tests (Tools)

describe('readFile tool', () => {
  it('reads file content correctly', async () => {
    const result = await readFile({ path: '/test/file.txt' });
    expect(result.success).toBe(true);
    expect(result.data).toContain('expected content');
  });
});

Behavior Tests (Agent Decisions)

describe('agent planning', () => {
  it('creates plan before executing file modifications', async () => {
    const trace = await agent.runWithTrace('Refactor the auth module');

    // Verify planning happened first
    const firstToolCall = trace.toolCalls[0];
    expect(firstToolCall.name).toBe('read_file');

    // Verify no writes without reads
    const writeIndex = trace.toolCalls.findIndex(t => t.name === 'write_file');
    const readIndex = trace.toolCalls.findIndex(t => t.name === 'read_file');
    expect(readIndex).toBeLessThan(writeIndex);
  });
});

Evaluation Tests

// Run nightly, not in regular CI
describe('Agent Accuracy (Eval)', () => {
  const testCases = loadTestCases('./evals/coding-tasks.json');

  it.each(testCases)('completes $name correctly', async (testCase) => {
    const result = await agent.run(testCase.input);

    // Verify against expected outcomes
    expect(result.filesModified).toEqual(testCase.expectedFiles);
    expect(await runTests(testCase.testCommand)).toBe(true);
  }, 120000);
});

Pydantic AI Patterns (Python Default)

Project Structure (Python)

project/
├── src/
│   ├── agents/
│   │   ├── __init__.py
│   │   ├── researcher.py       # Research agent
│   │   ├── coder.py            # Coding agent
│   │   └── orchestrator.py     # Main coordinator
│   ├── tools/
│   │   ├── __init__.py
│   │   ├── web.py              # Web search tools
│   │   ├── files.py            # File operations
│   │   └── database.py         # DB queries
│   ├── models/
│   │   ├── __init__.py
│   │   └── schemas.py          # Pydantic models
│   └── deps.py                 # Dependencies
├── tests/
│   ├── test_agents.py
│   └── test_tools.py
└── pyproject.toml

Agent with Tools

from pydantic_ai import Agent, RunContext
from pydantic import BaseModel
from httpx import AsyncClient

class SearchResult(BaseModel):
    title: str
    url: str
    snippet: str

class ResearchDeps(BaseModel):
    http_client: AsyncClient
    api_key: str

research_agent = Agent(
    'claude-sonnet-4-20250514',
    deps_type=ResearchDeps,
    result_type=list[SearchResult],
    system_prompt='You are a research assistant. Use tools to find information.',
)

@research_agent.tool
async def web_search(ctx: RunContext[ResearchDeps], query: str) -> list[dict]:
    """Search the web for information."""
    response = await ctx.deps.http_client.get(
        'https://api.search.com/search',
        params={'q': query},
        headers={'Authorization': f'Bearer {ctx.deps.api_key}'},
    )
    return response.json()['results']

@research_agent.tool
async def read_webpage(ctx: RunContext[ResearchDeps], url: str) -> str:
    """Read and extract content from a webpage."""
    response = await ctx.deps.http_client.get(url)
    return response.text[:5000]  # Truncate for context

# Usage
async def main():
    async with AsyncClient() as client:
        deps = ResearchDeps(http_client=client, api_key='...')
        result = await research_agent.run(
            'Find recent articles about LLM agents',
            deps=deps,
        )
        for item in result.data:
            print(f"- {item.title}")

Structured Output with Validation

from pydantic import BaseModel, Field
from pydantic_ai import Agent

class CodeReview(BaseModel):
    summary: str = Field(description="Brief summary of the review")
    issues: list[str] = Field(description="List of issues found")
    suggestions: list[str] = Field(description="Improvement suggestions")
    approval: bool = Field(description="Whether code is approved")
    confidence: float = Field(ge=0, le=1, description="Confidence score")

review_agent = Agent(
    'claude-sonnet-4-20250514',
    result_type=CodeReview,
    system_prompt='Review code for quality, security, and best practices.',
)

# Result is validated Pydantic model
result = await review_agent.run(f"Review this code:\n```python\n{code}\n```")
if result.data.approval:
    print("Code approved!")
else:
    for issue in result.data.issues:
        print(f"Issue: {issue}")

Multi-Agent Coordination

from pydantic_ai import Agent

# Specialized agents
planner = Agent('claude-sonnet-4-20250514', system_prompt='Create detailed plans.')
executor = Agent('claude-sonnet-4-20250514', system_prompt='Execute tasks precisely.')
reviewer = Agent('claude-sonnet-4-20250514', system_prompt='Review and verify work.')

async def orchestrate(task: str):
    # 1. Plan
    plan = await planner.run(f"Create a plan for: {task}")

    # 2. Execute each step
    results = []
    for step in plan.data.steps:
        result = await executor.run(f"Execute: {step}")
        results.append(result.data)

    # 3. Review
    review = await reviewer.run(
        f"Review the results:\nTask: {task}\nResults: {results}"
    )

    return review.data

Streaming Responses

from pydantic_ai import Agent

agent = Agent('claude-sonnet-4-20250514')

async def stream_response(prompt: str):
    async with agent.run_stream(prompt) as response:
        async for chunk in response.stream():
            print(chunk, end='', flush=True)

    # Get final structured result
    result = await response.get_data()
    return result

Testing Agents

import pytest
from pydantic_ai import Agent
from pydantic_ai.models.test import TestModel

@pytest.fixture
def test_agent():
    return Agent(
        TestModel(),  # Mock model for testing
        result_type=str,
    )

async def test_agent_response(test_agent):
    result = await test_agent.run('Test prompt')
    assert result.data is not None

# Test with specific responses
async def test_with_mock_response():
    model = TestModel()
    model.seed_response('Expected output')

    agent = Agent(model)
    result = await agent.run('Any prompt')
    assert result.data == 'Expected output'

Skills Pattern (Anthropic)

Skill Structure

skills/
└── code-review/
    ├── instructions.md      # How to perform code reviews
    ├── scripts/
    │   └── run-linters.sh   # Supporting scripts
    └── resources/
        └── checklist.md     # Review checklist

instructions.md Example

# Code Review Skill

## When to Use
Activate this skill when asked to review code, PRs, or diffs.

## Process
1. Read the changed files completely
2. Run linters: `./scripts/run-linters.sh`
3. Check against resources/checklist.md
4. Provide structured feedback

## Output Format
- Summary (1-2 sentences)
- Issues found (severity: critical/major/minor)
- Suggestions for improvement
- Approval recommendation

Loading Skills Dynamically

async function loadSkill(skillName: string): Promise<Skill> {
  const skillPath = `./skills/${skillName}`;
  const instructions = await fs.readFile(`${skillPath}/instructions.md`, 'utf-8');
  const scripts = await glob(`${skillPath}/scripts/*`);
  const resources = await glob(`${skillPath}/resources/*`);

  return {
    name: skillName,
    instructions,
    scripts: scripts.map(s => ({ name: path.basename(s), path: s })),
    resources: await Promise.all(resources.map(loadResource)),
  };
}

Anti-Patterns

• No planning before execution - Agents that jump to action make more errors
• Monolithic agents - One agent with 50 tools becomes confused
• No verification - Agents must verify their own work
• Hardcoded tool sequences - Let the model decide tool order
• Missing guardrails - All agents need safety boundaries
• No state management - Lose context across tool calls
• Testing only happy paths - Test failures and edge cases
• Ignoring model differences - Reasoning models need different prompts
• No cost tracking - Agentic workflows can be expensive
• Full automation without oversight - Human-in-the-loop for critical actions

Quick Reference

Agent Development Checklist

• Define clear agent scope and boundaries
• Design tools with explicit schemas and risk levels
• Implement explore-plan-execute-verify workflow
• Add multi-layer guardrails
• Set up conversation and persistent memory
• Write behavior and evaluation tests
• Configure appropriate model for task complexity
• Add human-in-the-loop for high-risk operations
• Monitor token usage and costs
• Document skills and instructions

Thinking Triggers (Claude)

"think"        → Standard analysis
"think hard"   → Deeper reasoning
"think harder" → Extensive analysis
"ultrathink"   → Maximum reasoning

Gemini Settings

thinking_level: "high" | "low"
temperature: 1.0 (keep at 1.0 for reasoning)
thoughtSignature: <pass back for function calling>