openai-agents-sdk

The openai-agents-sdk skill provides comprehensive assistance for developers building sophisticated autonomous agents using the Python openai-agents package. It is designed for engineers and AI researchers tasked with orchestrating complex multi-agent workflows, implementing strict Pydantic-based output validation, and managing persistent conversation sessions. Whether you are building internal automation tools, research assistants, or complex multi-step reasoning pipelines, this skill bridges the gap between raw API implementation and robust production-grade agent architecture.

• Full lifecycle management for OpenAI agents including Agent initialization, instruction crafting, and model selection for gpt-5 series.

• Deep integration for multi-agent handoffs, allowing specialized agents to delegate tasks and coordinate complex business logic.

• Support for custom function tools via @function_tool decorators and AgentOutputSchema for reliable JSON structured output extraction.

• Advanced execution controls including synchronous and asynchronous runs (Runner.run_sync, Runner.run_streamed) and stateful SQLiteSession handling.

• Native support for Azure OpenAI and multi-provider backends configured through LiteLLM, ensuring seamless migration and model flexibility.

• Built-in observability patterns for execution graph recording, guardrails for input/output safety, and sandboxed execution environments.

• Live documentation retrieval via MCP (Model Context Protocol) to ensure model IDs and API signatures remain accurate during fast-paced development cycles.

• Best practices: Always use the openaiDeveloperDocs MCP server to verify model IDs and API schemas before finalizing code, as OpenAI SDK updates are frequent.

• Usage: Ensure your environment variables (OPENAI_API_KEY, LLM_PROVIDER, AZURE_API_KEY) are correctly configured before execution.

• Limitations: While designed for the agents SDK, complex multi-agent orchestration requires careful design of handoff triggers and context isolation.

• Keywords: Python agents, OpenAI SDK, multi-agent pipeline, function tools, AgentOutputSchema, SQLiteSession, LiteLLM, guardrails, LLM as judge, streaming, orchestration, Pydantic validation.