Build an Agent Factory with OpenAI’s Agents API and Rasa Mantle

Build an Agent Factory with OpenAI’s Agents API and Rasa Mantle Agents that build, test, and improve other agents. A hands-on workshop for AI builders and product teams. You already use AI to write code. What happens when you give it the tools to test an agent, investigate failures, and propose the next improvement? In this Rasa workshop, we’ll connect OpenAI’s Agents API with Rasa Mantle to build a small agent factory: a workflow where a cloud agent helps improve a customer-facing AI agent, and your team reviews the results. Start with a working voice agent. Challenge it with a difficult customer request or a failed business-system call. Then put a cloud agent to work investigating what happened, proposing a fix and rerunning the checks. You’ll inspect the conversation, the outlined changes and the before-and-after results—and decide whether the new version is ready to release. What you’ll build Working in small groups with a prepared starter project, you’ll: Run a Rasa Mantle customer-service agent connected to a simulated order system. Use OpenAI’s Agents API to investigate a failure and propose an improvement. Connect the development workflow to Rasa’s simulation and evaluation tools through MCP. Test the proposed change against the initial failure and existing scenarios. Review the evidence before approving the next version. The customer scenario is deliberately small: checking an order and requesting a delivery-address change. The interesting part is the factory around it—how you turn a failure into a tested improvement. What you’ll take away A runnable starter project, an example improvement workflow, before-and-after evaluation results, and a practical pattern for deciding what to automate next. We’ll also discuss how this pattern extends to production: turning observed failures into reproducible tests, controlling what agents can change, and keeping release decisions accountable. Who should attend AI engineers, developers, technical founders, and product le