Skip to main content

​
Overview

The general agent is the primary AI assistant for students, capable of web search, code execution, Canvas LMS integration, and memory management.

​
System Prompt Structure

The system prompt is built dynamically for each request: 
export function buildSystemPrompt(ctx: AgentContext): string {
  return `You are a general assistant and coordinator for students at ${ctx.schoolName}.

🔍 YOUR CAPABILITIES:
- **Web search**: Search for current information beyond your knowledge cutoff
- **Code execution**: Run Python code for calculations, data processing, and analysis
- **Programmatic tool calling**: Chain multiple tool calls efficiently within code
- **Document processing**: Create and edit PowerPoint, Word, Excel, and PDF files
- **Canvas LMS search**: Search student's course content, assignments, and materials
- **Memory**: Remember important information about the user across conversations

🎯 CURRENT CONTEXT:
- **Date/Time**: ${ctx.currentDateTime} (${ctx.timezone})
- **School**: ${ctx.schoolName}
- **User**: ${ctx.fullName}
...`;
}

​
AgentContext Interface

The context object passed to all agent functions:
​
AgentContext
interface

​
Tool Configuration

Tools are configured with provider-specific options: 
export function getGeneralAgentTools(ctx: AgentContext): Record<string, any> {
  return {
    // Code execution with programmatic tool calling
    code_execution: anthropic.tools.codeExecution_20250825(),

    // Memory tool with filesystem-like interface
    memory: anthropic.tools.memory_20250818({
      execute: async (action) => {
        return await executeMemoryCommand(ctx.userId, action);
      },
    }),

    // Location-aware web search
    web_search: anthropic.tools.webSearch_20250305({
      userLocation: {
        type: "approximate",
        city: ctx.city,
        country: ctx.country,
        region: ctx.region,
        timezone: ctx.timezone,
      },
    }),

    web_fetch: anthropic.tools.webFetch_20250910(),

    // Canvas LMS tools
    searchContent: searchContentTool,
    getClassAssignments: getClassAssignmentsTool,

    // Todo management
    getTodos: getTodosTool,
    createTodo: createTodoTool,
    updateTodo: updateTodoTool,
    deleteTodo: deleteTodoTool,

    // Study tools
    createStudySet: createStudySetTool,
  };
}

​
Programmatic Tool Calling Pattern

The agent is instructed to use Python for complex workflows: 
import json
import asyncio

# Fetch all assignments
result = await getClassAssignments({})
assignments = json.loads(result)

# Filter and process
overdue = [a for a in assignments if is_overdue(a['due_at'])]
print(f"Found {len(overdue)} overdue assignments")
Web search automatically uses the user’s location: 
web_search: anthropic.tools.webSearch_20250305({
  userLocation: {
    type: "approximate",
    city: ctx.city,        // e.g., "Boston"
    country: ctx.country,  // e.g., "USA"
    region: ctx.region,    // e.g., "Massachusetts"
    timezone: ctx.timezone // e.g., "America/New_York"
  }
})
This enables queries like “best coffee shops near me” to return location-relevant results.

​
Style Guidelines

From the system prompt:
  • Default style is natural, chatty, and playful, not formal
  • Keep tone topic-appropriate and matched to the user
  • Brief responses for casual conversation
  • Use emojis and casual punctuation if user does
  • Only in prose (not in section headers)
  • Don’t use Markdown sections/lists in casual chat unless asked
  • Use h1 (#) for section headers, not bold
  • Format structured responses with Markdown
  • Use LaTeX for math: $$x^2 + y^2 = z^2$$
  • Refer to classes by friendly name only (IDs are internal)
  • Convert technical formats (snake_case, UUIDs, paths) to plain language
  • Keep content appropriate for ages 13-18
  • Never say “I don’t have access to the internet” - use web_search!
  • Use tools proactively without asking permission
  • Check memory first to avoid asking redundant questions
  • Proactively save important facts to memory
  • Convert developer formats to plain language before answering

​
Skills Available

Documentation loaded into the code execution container:

canvas-assignments

Canvas assignment data structure, fields, date handling, common patterns

todo-management

Todo data structure, views, date types, Canvas linking, CRUD operations

​
User Classes Context

The agent receives a formatted list of the user’s classes: 
ctx.classes.map((course) => `- ${course.name} (ID: ${course.id})`).join("\n")
Example output: 
📚 USER'S CLASSES:
- Calculus I (ID: 12345)
- Introduction to Psychology (ID: 67890)
- English Composition (ID: 11111)

​
Implementation Reference

// apps/web/src/ai/agents/general.ts
export interface AgentContext {
  userId: string;
  fullName: string;
  schoolName: string;
  classes: CanvasCourse[];
  currentDateTime: string;
  timezone: string;
  chatId: string;
  country?: string;
  city?: string;
  region?: string;
}

export function buildSystemPrompt(ctx: AgentContext): string {
  // Full system prompt with capabilities, tools, and guidelines
}

export function getGeneralAgentTools(ctx: AgentContext): Record<string, any> {
  // Tool configuration with provider options
}

​
Next Steps

Available Tools

Explore all Canvas, Todo, and Study tools

Memory System

Learn about persistent user memory

Build docs developers (and LLMs) love

Get started for freeTalk to us