Literature: OpenAI Agents SDK

Overview

The official documentation for the OpenAI Agents SDK provides a comprehensive guide to building production-ready agentic applications. It supersedes the experimental openai-swarm with a focus on durability, safety, and multi-agent orchestration.

Key Concepts

Agents and Handoffs

• Agents are LLMs equipped with instructions, tools, and handoffs.
• Handoffs are a specific type of tool that transfers control from one agent to another.
• Control is managed by a "manager" agent or handled via direct handoffs where the specialist becomes the active agent.

Orchestration Patterns

• Orchestrating via LLM: Agents autonomously plan and delegate using tools and handoffs.
• Orchestrating via Code: Developers determine the flow of agents through imperative Python logic.
• Mixed patterns allow for flexible yet controlled workflows.

Model Context Protocol (MCP)

• The SDK provides native support for MCP.
• MCPServer base class for building custom servers.
• Integration allows agents to use any MCP-compliant tool or resource.

Advanced Features

• Guardrails: Input and output validation to ensure safety and correctness.
• Tracing: Built-in visualization and debugging of agentic flows.
• Sandbox Agents: Execution in isolated workspaces with persistent files and sessions.
• Human-in-the-Loop: Integrated patterns for manual approval of agent actions.

Ingestion Details

• Date Ingested: 2026-05-01
• Tool: Firecrawl Pipeline (Scrape/Crawl)
• Pages Processed: 15
• Status: Indexed in Supabase (pgvector)

Connections

• openai-agents-sdk (Permanent Note)
• mcp-moc
• agent-thought-cycle
• pattern-human-in-the-loop