backend-ai-agent（后端 AIAgent）

Backend AI Agent Creation

Create AI agents using the Vercel AI SDK with proper patterns for tool use, tracing, and error handling.

Overview

The AI SDK supports:

• One-off calls: generateText and generateObject for single-purpose operations
• Agentic workflows: Multi-turn conversations with tool use

Installation

npm install ai @ai-sdk/anthropic @ai-sdk/google @ai-sdk/openai zod

Setup

1. Model Providers (lib/ai/models.ts)

import { createAnthropic } from '@ai-sdk/anthropic';
import { createGoogleGenerativeAI } from '@ai-sdk/google';
import { createOpenAI } from '@ai-sdk/openai';
import config from '../../config';

export const gemini = createGoogleGenerativeAI({ apiKey: config.gemini.apiKey });
export const claude = createAnthropic({ apiKey: config.anthropic.apiKey });
export const openai = createOpenAI({ apiKey: config.openai.apiKey });

2. Model Constants (in @{project}/types)

export const LATEST_PRO_CLAUDE_MODEL = 'claude-sonnet-4-20250514';
export const LATEST_PRO_OPENAI_MODEL = 'gpt-4o-2025-04-15';
export const LATEST_PRO_GEMINI_MODEL = 'gemini-2.0-flash-exp';
export const CLAUDE_HAIKU_MODEL = 'claude-haiku-4.5-20250103';
export const GEMINI_FLASH_MODEL = 'gemini-2.0-flash-exp';

export type ModelProvider = 'openai' | 'claude' | 'gemini';

3. Config

Add to apps/backend/src/config/index.ts:

anthropic: { apiKey: process.env.ANTHROPIC_API_KEY || '' },
openai: { apiKey: process.env.OPENAI_API_KEY || '' },
gemini: { apiKey: process.env.GEMINI_API_KEY || '' },

One-Off LLM Calls

Text Generation

import { generateText } from 'ai';
import { gemini } from '../lib/ai/models';
import { GEMINI_FLASH_MODEL } from '@{project}/types';

export async function generateSummary(content: string): Promise<string | null> {
  try {
    const result = await generateText({
      model: gemini(GEMINI_FLASH_MODEL),
      prompt: `Summarize the following content:\n\n${content}`,
    });
    return result.text;
  } catch (err) {
    log.error({ err }, 'Error generating summary');
    return null;
  }
}

Structured Output

import { generateObject } from 'ai';
import { z } from 'zod/v4';

const TimeEstimateSchema = z.object({
  hours: z.number().int().min(0).describe('Estimated hours'),
  minutes: z.number().int().min(0).max(59).describe('Additional minutes'),
  reasoning: z.string().describe('Explanation'),
});

const result = await generateObject({
  model: gemini(GEMINI_FLASH_MODEL),
  schema: TimeEstimateSchema,
  prompt: `Estimate how long this task will take: ${task}`,
});

Agentic Workflows

See references/agent-patterns.md for complete examples including:

• Basic agent structure with conversation loop
• Tool creation patterns
• Model failover implementation

Model Selection Guide

Best Practices

Prompt Structure

const prompt = `
  You are an expert at [specific task].

  <context>
  ${relevantContext}
  </context>

  <instructions>
  1. [First instruction]
  2. [Second instruction]
  </instructions>
`;

Schema Design

Make schemas descriptive:

const schema = z.object({
  category: z.enum(['urgent', 'normal', 'low'])
    .describe('Priority level based on deadline and impact'),
  confidence: z.number().min(0).max(1)
    .describe('Confidence score from 0 to 1'),
});

Error Handling

Always handle errors gracefully with safe defaults:

try {
  const result = await generateObject({ model, schema, prompt });
  return result.object;
} catch (error) {
  log.error({ error }, 'AI call failed');
  return { category: 'normal', confidence: 0, reasoning: 'Error occurred' };
}

Token Limits

• One-off calls: maxOutputTokens: 2048
• Agentic workflows: maxOutputTokens: 8192
• Complex structured output: maxOutputTokens: 16384

File Structure

apps/backend/src/lib/ai/
├── models.ts           # Model provider instances
├── failover.ts         # Failover utilities
├── agents/
│   └── MyAgent.ts      # Agent implementations
└── tools/
    └── myTools.ts      # Tool definitions

Checklist

1. Install dependencies: npm install ai @ai-sdk/anthropic @ai-sdk/google @ai-sdk/openai
2. Create lib/ai/models.ts with provider instances
3. Add model constants to @{project}/types
4. Add API keys to config
5. Create agent class with conversation loop
6. Create tools with Zod parameters
7. Implement failover for critical operations
8. Test with realistic inputs