AI Framework Guides & Integrations

Relevant source files

The following files were used as context for generating this wiki page:

• guides/typescript/mastra/coding-agent/openai/src/mastra/agents/coding-agent.ts
• guides/typescript/mastra/coding-agent/openai/README.md
• guides/typescript/anthropic/multi-agent-claude-sdk/README.md
• guides/typescript/agentkit-inngest/coding-agent/anthropic/README.md
• guides/python/recursive-language-models/rlm/prompts.py
• guides/python/ai-data-analyst/openai/README.md
• guides/typescript/openai/codex-sdk/README.md

The AI Framework Guides & Integrations section provides a comprehensive set of implementations and architectural patterns for deploying AI agents within secure Daytona sandboxes. These integrations enable LLMs to perform complex software engineering, data analysis, and multi-agent coordination tasks by leveraging Daytona's isolated execution environments.

The primary purpose of these integrations is to bridge the gap between AI reasoning and physical code execution. By providing specialized toolsets for sandbox management, file operations, and shell execution, Daytona allows frameworks like Mastra, AgentKit, and the Claude Agent SDK to interact safely with real-world development environments.

Sources: guides/typescript/mastra/coding-agent/openai/README.md, guides/typescript/anthropic/multi-agent-claude-sdk/README.md

Core Integration Patterns

Multi-Agent Hierarchical Architecture

The project demonstrates a hierarchical "Two-Agent" system designed to separate high-level planning from low-level execution. In this architecture, a Project Manager Agent runs locally to orchestrate tasks, while a Developer Agent operates within a Daytona sandbox.

The Project Manager uses specialized XML-like tags (<developer_task>) to delegate work, and the Developer Agent uses the Claude Agent SDK within the sandbox to perform file operations and start services.

Sources: guides/typescript/anthropic/multi-agent-claude-sdk/README.md

Tool-Based Agentic Frameworks (Mastra & AgentKit)

Integrations with frameworks like Mastra and Inngest's AgentKit focus on providing the agent with a "Sandbox Toolkit." This toolkit includes atomic operations that the agent can call to manage its environment.

Tool Category	Key Functions	Purpose
Sandbox Mgmt	createSandbox	Initialize isolated execution environments
Execution	runCode, runCommand	Execute Python/JS/TS or shell commands with real-time output
File Ops	writeFile, writeFiles, readFile	CRUD operations and batch file creation
Monitoring	watchDirectory, getFileInfo	Live tracking of filesystem changes and metadata

Sources: guides/typescript/mastra/coding-agent/openai/src/mastra/agents/coding-agent.ts:25-56, guides/typescript/mastra/coding-agent/openai/README.md

Specialized AI Workflows

Recursive Language Model (RLM) Agents

The RLM integration uses a recursive spawning pattern where a parent agent can create sub-agents to handle complex sub-tasks in parallel. Each sub-agent receives a "fresh" sandbox copy to prevent environmental pollution during investigation.

The RLM system enforces strict formatting, requiring a THOUGHT section and exactly one Python block per response to ensure predictable execution.

Sources: guides/python/recursive-language-models/rlm/prompts.py:7-124

AI Data Analyst Workflow

The Data Analyst integration automates the process of data ingestion, processing, and visualization. It utilizes a two-step LLM process: a coding model for high-accuracy script generation and a summary model for human-readable reporting.

1. Code Generation: LLM generates Python code based on a CSV data format and natural language prompt.
2. Sandbox Setup: A Daytona sandbox is initialized, and the dataset is uploaded.
3. Execution: The generated script runs, typically using libraries like Pandas for analysis and Matplotlib for chart generation.
4. Result Retrieval: Generated artifacts (e.g., chart-0.png) are downloaded from the sandbox to the host.

Sources: guides/python/ai-data-analyst/openai/README.md, guides/typescript/ai-data-analyst/openai/README.md

Implementation Details

Mastra Coding Agent Configuration

The Mastra agent is configured with a persistent memory system using LibSQL and semantic recall capabilities.

export const codingAgent = new Agent({
  id: 'coding-agent',
  name: 'Coding Agent',
  instructions: `...`,
  model: process.env.MODEL || 'openai/gpt-4.1',
  tools: {
    createSandbox,
    runCode,
    readFile,
    writeFile,
    // ... additional tools
  },
  memory: new Memory({
    storage: new LibSQLStore({ id: 'coding-agent-storage', url: 'file:../../mastra.db' }),
    options: {
      generateTitle: true,
      semanticRecall: true,
    },
  }),
})

Sources: guides/typescript/mastra/coding-agent/openai/src/mastra/agents/coding-agent.ts:167-195

Environment Configuration

Integrations across different frameworks share a common set of required environment variables for authentication and model routing.

Variable	Required	Description
DAYTONA_API_KEY	Yes	Authentication for the Daytona API and sandbox creation
OPENAI_API_KEY	Optional	Key for OpenAI models (GPT-4o, Codex)
ANTHROPIC_API_KEY	Optional	Key for Anthropic models (Claude 3.5 Sonnet/Haiku)
SANDBOX_OPENAI_API_KEY	Optional	Specific key passed into the sandbox environment for agent use

Sources: guides/typescript/openai/codex-sdk/README.md, guides/typescript/agentkit-inngest/coding-agent/anthropic/README.md

Summary

The AI Framework Guides & Integrations module establishes a standardized way to run AI-generated code securely. By utilizing Daytona's infrastructure, developers can build autonomous agents that plan, write, and execute code with full access to a professional development toolkit, including multi-language support (Python, JS, TS) and real-time monitoring. These patterns enable sophisticated workflows ranging from parallelized codebase exploration to automated visual data analysis.