Combining Agents SDK and MCP: Building Powerful Multi-Agent Workflows With Ease

Combining Agents SDK and MCP: Building Powerful Multi-Agent Workflows With Ease

Why we build creative tech for real people.

✨ Introduction

In the rapidly evolving world of AI development, intelligent agents are becoming more than just a novelty—they’re a foundation for how we build with large language models. As agents get smarter, they need structured, reliable access to real-world tools.

Over the weekend, I built a project that leverages two key players: OpenAI’s Agents SDK and Anthropic’s Model Context Protocol (MCP).

This blog post walks through how I combined the Agents SDK with MCP to build a flexible, modular multi-agent system—and why this architecture matters for the future of AI-native apps.

🛠 What is the Agents SDK?

OpenAI’s Agents SDK is a developer toolkit for creating AI agents that can reason, plan, and use tools.

Key features include:

  • Turn-based interaction – Agents process one message at a time and decide what to do.
  • Tool calling – Agents can call external APIs through function calls.
  • Stateful memory – Agents can remember previous interactions.
  • Agent handoffs – One agent can delegate tasks to another.

This makes it ideal for building multi-step workflows, customer support bots, or even internal copilots.

🌐 What is MCP?

MCP (Model Context Protocol) is an open standard introduced by Anthropic. It defines how assistants communicate with external tools in a structured way.

MCP servers handle:

  • Tool registration – Each service describes its capabilities using JSON Schema.
  • Request handling – The server listens for function calls.
  • Structured output – Responses are returned in machine-readable form.

MCP implementations like mcp-notion-server, mcp-slack-server, and mcp-github-server make it easy to integrate Notion, Slack, and GitHub into LLM-powered agents.

🤝 OpenAI x Anthropic? Yes, Really

In a surprising but welcome move, OpenAI announced on March 27th, 2025 that the Agents SDK now supports MCP—the protocol originally launched by Anthropic, a direct competitor.

This is huge.

It shows OpenAI’s commitment to interoperability and the open-source ecosystem. By embracing MCP, OpenAI is making it easier for developers to build rich, agentic workflows across tools, no matter where they come from.

🔗 Why Combine Them?

Combining the Agents SDK and MCP gives you the best of both worlds:

Agents SDK

  • Handles orchestration, reasoning, and handoffs
  • Powered by OpenAI models
  • Modular, turn-based design
  • JSON schema-defined capabilities
  • Easy debugging through tracing dashboard

MCP

  • Bridges agents with real-world tools
  • Tool-agnostic, open standard

When combined, you can build workflows like:

“Summarize my Notion tasks and post them to Slack.”

And under the hood, this happens:

  1. 🧠 A Triage Agent receives the request.
  2. 🗂 It hands off to the Notion Agent.
  3. 📝 Notion Agent fetches and summarizes tasks (via mcp-notion-server).
  4. 📬 A second handoff sends the summary to the Slack Agent to post it (via mcp-slack-server).

Each sub-agent operates independently using MCPServerStdio, which links them to their respective external tools via MCP. The Triage Agent employs the handoff mechanism from the Agents SDK to route user requests to the appropriate sub-agent based on intent.

This architecture is modular, scalable, and easy to extend.

🐛 Debugging & Contributing

While testing, I kept hitting errors from OpenAI’s Responses API on one of the MCP servers.

Turns out:

  • Tools properties were missing an items schema.

So I fixed it, added proper JSON schema definitions, and submitted a pull request to the mcp-notion-server repo.

💡 Open source contributions like this help move the entire ecosystem forward—by fixing bugs, improving dev experience, and making tools more reliable for everyone.

💬 Final Thoughts

The future of agent development is collaborative, modular, and open.

By combining OpenAI’s Agents SDK with Anthropic’s Model Context Protocol, you can build robust multi-agent systems that interact across tools with ease.

This integration also reflects a deeper shift: openness and interoperability are becoming the norm in AI development.

I’m excited to keep building with this stack—and I hope you’ll try it too.