convex-agents

# Convex Agents

Build persistent, stateful AI agents with Convex including thread management, tool integration, streaming responses, RAG patterns, and workflow orchestration.

## Documentation Sources

Before implementing, do not assume; fetch the latest documentation:

- Primary: https://docs.convex.dev/ai
- Convex Agent Component: https://www.npmjs.com/package/@convex-dev/agent
- For broader context: https://docs.convex.dev/llms.txt

## Instructions

### Why Convex for AI Agents

- **Persistent State** - Conversation history survives restarts
- **Real-time Updates** - Stream responses to clients automatically
- **Tool Execution** - Run Convex functions as agent tools
- **Durable Workflows** - Long-running agent tasks with reliability
- **Built-in RAG** - Vector search for knowledge retrieval

### Setting Up Convex Agent

```bash
npm install @convex-dev/agent ai openai
```

```typescript
// convex/agent.ts
import { Agent } from "@convex-dev/agent";
import { components } from "./_generated/api";
import { OpenAI } from "openai";

const openai = new OpenAI();

export const agent = new Agent(components.agent, {
  chat: openai.chat,
  textEmbedding: openai.embeddings,
});
```

### Thread Management

```typescript
// convex/threads.ts
import { mutation, query } from "./_generated/server";
import { v } from "convex/values";
import { agent } from "./agent";

// Create a new conversation thread
export const createThread = mutation({
  args: {
    userId: v.id("users"),
    title: v.optional(v.string()),
  },
  returns: v.id("threads"),
  handler: async (ctx, args) => {
    const threadId = await agent.createThread(ctx, {
      userId: args.userId,
      metadata: {
        title: args.title ?? "New Conversation",
        createdAt: Date.now(),
      },
    });
    return threadId;
  },
});

// List user's threads
export const listThreads = query({
  args: { userId: v.id("users") },
  returns: v.array(v.object({
    _id: v.id("threads"),
    title: v.string(),
    lastMessageAt: v.optional(v.number()),
  })),
  handler: async (ctx, args) => {
    return await agent.listThreads(ctx, {
      userId: args.userId,
    });
  },
});

// Get thread messages
export const getMessages = query({
  args: { threadId: v.id("threads") },
  returns: v.array(v.object({
    role: v.string(),
    content: v.string(),
    createdAt: v.number(),
  })),
  handler: async (ctx, args) => {
    return await agent.getMessages(ctx, {
      threadId: args.threadId,
    });
  },
});
```

### Sending Messages and Streaming Responses

```typescript
// convex/chat.ts
import { action } from "./_generated/server";
import { v } from "convex/values";
import { agent } from "./agent";
import { internal } from "./_generated/api";

export const sendMessage = action({
  args: {
    threadId: v.id("threads"),
    message: v.string(),
  },
  returns: v.null(),
  handler: async (ctx, args) => {
    // Add user message to thread
    await ctx.runMutation(internal.chat.addUserMessage, {
      threadId: args.threadId,
      content: args.message,
    });

    // Generate AI response with streaming
    const response = await agent.chat(ctx, {
      threadId: args.threadId,
      messages: [{ role: "user", content: args.message }],
      stream: true,
      onToken: async (token) => {
        // Stream tokens to client via mutation
        await ctx.runMutation(internal.chat.appendToken, {
          threadId: args.threadId,
          token,
        });
      },
    });

    // Save complete response
    await ctx.runMutation(internal.chat.saveResponse, {
      threadId: args.threadId,
      content: response.content,
    });

    return null;
  },
});
```

### Tool Integration

Define tools that agents can use:

```typescript
// convex/tools.ts
import { tool } from "@convex-dev/agent";
import { v } from "convex/values";
import { api } from "./_generated/api";

// Tool to search knowledge base
export const searchKnowledge = tool({
  name: "search_knowledge",
  description: "Search the knowledge base for relevant information",
  parameters: v.object({
    query: v.string(),
    limit: v.optional(v.number()),
  }),
  handler: async (ctx, args) => {
    const results = await ctx.runQuery(api.knowledge.search, {
      query: args.query,
      limit: args.limit ?? 5,
    });
    return results;
  },
});

// Tool to create a task
export const createTask = tool({
  name: "create_task",
  description: "Create a new task for the user",
  parameters: v.object({
    title: v.string(),
    description: v.optional(v.string()),
    dueDate: v.optional(v.string()),
  }),
  handler: async (ctx, args) => {
    const taskId = await ctx.runMutation(api.tasks.create, {
      title: args.title,
      description: args.description,
      dueDate: args.dueDate ? new Date(args.dueDate).getTime() : undefined,
    });
    return { success: true, taskId };
  },
});

// Tool to get weather
export const getWeather = tool({
  name: "get_weather",
  description: "Get current weather for a location",
  parameters: v.object({
    location: v.string(),
  }),
  handler: async (ctx, args) => {
    const response = await fetch(
      `https://api.weather.com/current?location=${encodeURIComponent(args.location)}`
    );
    return await response.json();
  },
});
```

### Agent with Tools

```typescript
// convex/assistant.ts
import { action } from "./_generated/server";
import { v } from "convex/values";
import { agent } from "./agent";
import { searchKnowledge, createTask, getWeather } from "./tools";

export const chat = action({
  args: {
    threadId: v.id("threads"),
    message: v.string(),
  },
  returns: v.string(),
  handler: async (ctx, args) => {
    const response = await agent.chat(ctx, {
      threadId: args.threadId,
      messages: [{ role: "user", content: args.message }],
      tools: [searchKnowledge, createTask, getWeather],
      systemPrompt: `You are a helpful assistant. You have access to tools to:
        - Search the knowledge base for information