OpenAI AgentKit Expert Skill

Purpose

This skill provides comprehensive guidance on building production-ready multi-agent systems using OpenAI's AgentKit platform and Agents SDK, following 2025 best practices.

Platform Overview

OpenAI AgentKit (2025)

Complete platform for building, deploying, and optimizing agents with enterprise-grade tooling.

Core Components:

• Agent Builder - Visual canvas for creating and versioning multi-agent workflows
• Connector Registry - Central management for data and tool connections
• ChatKit - Embeddable customizable chat-based agent experiences
• Evaluation Suite - Datasets, trace grading, automated prompt optimization
• Multi-Model Support - Third-party model integration capabilities

Agents SDK (Production-Ready)

The Agents SDK is the production evolution of the experimental Swarm framework. Use Agents SDK for all production work - Swarm is educational only.

Migration Note: If you encounter legacy Swarm code, migrate to Agents SDK immediately.

Core Concepts

1. Agents

An Agent encapsulates:

• A set of instructions (system prompt)
• A set of functions/tools
• The capability to hand off execution to another Agent

Design Principle: Agents should be lightweight and specialized rather than monolithic and general-purpose.

2. Routines

A routine is a sequence of actions an agent can perform:

• Natural language instructions (via system prompt)
• Available tools needed to execute
• Context and state management
• Success criteria

Think of it as: A mini-workflow that an agent owns and executes autonomously.

3. Handoffs

Handoffs enable agent-to-agent transitions in execution flow.

Key Pattern: When an agent encounters a task outside its specialization, it hands off to a more appropriate agent.

Example:

# Triage agent determines which specialist to use
if task.type == "refund":
    handoff_to(refund_agent)
elif task.type == "sales":
   ...