The Build Hour session, hosted by Sarah Urbonus and Alain from OpenAI, kicked off 2025 by focusing on agentic tool calling and how developers can leverage OpenAI’s latest APIs and models to build powerful, goal-oriented agents. They introduced the concept of agentic tool calling, which combines advanced reasoning capabilities with tool usage, enabling models to autonomously figure out the steps needed to achieve a desired outcome. This approach moves beyond step-by-step instructions, allowing models to learn from results and refine their strategies through reinforcement learning, making them robust and consistent over long, complex tasks.
Alain demonstrated how to implement an agentic task system using OpenAI’s agents SDK, showcasing a live coding session where they built an agent capable of handling customer service tickets by interacting with mock APIs for user data, order details, and refunds. The session highlighted how specifying the end goal rather than the steps allows the model to autonomously plan and execute tasks using available tools. They also introduced the concept of “to-dos” as a way for the agent to surface progress updates to users, enhancing transparency and user experience by showing task completion stages without overwhelming users with technical details.
The session then delved into backend architecture, illustrating how to set up a Flask or FastAPI server to manage tasks and stream events to the frontend using Server-Sent Events (SSE). This architecture supports both foreground tasks, where the connection remains open during task execution, and background tasks, which run asynchronously and allow users to disconnect and reconnect later. Alain explained how tasks are managed in the backend with asynchronous workers, event queues, and task objects that maintain state and progress, enabling scalable and responsive agent-driven applications.
Further, the discussion touched on delegation, where agents can autonomously start and manage subtasks in the background, allowing users to continue interacting with the system without waiting for long-running processes to complete. This non-blocking interaction model enhances user experience by enabling multitasking and real-time updates. Alain also answered audience questions on best practices for orchestrating tool calls, managing memory in long-horizon tasks, the ideal number of tools for an agent, and integrating OpenAI hosted functions with custom functions, providing practical guidance for developers building complex agentic systems.
The session concluded with resources for further learning, including GitHub repositories, previous build hours, and a practical guide for building agents. The hosts announced the next Build Hour focused on image generation APIs, encouraging developers to continue exploring and innovating with OpenAI’s evolving toolset. Overall, the session provided a comprehensive overview of building agentic systems that combine reasoning, tool usage, and task management to create sophisticated, autonomous AI applications.
