# Build Hour: Agentic Tool Calling Source: https://www.youtube.com/watch?v=7E-qdsVEoB8 Recap page: https://rapidrecap.app/video/7E-qdsVEoB8 Generated: 2025-09-03T01:01:38.027+00:00 --- ## Quick Overview OpenAI's Build Hour demonstrates how to leverage the Agents SDK and new APIs for agentic tool calling, enabling developers to build sophisticated, goal-oriented agents that can reason and perform complex, long-horizon tasks. The session highlights the practical implementation of these concepts through live coding, showcasing how to define agents with tools, integrate them into backend systems using Flask, and manage task execution with features like background processing and progress visualization via to-dos, ultimately empowering users to create more autonomous and capable AI applications. **Key Points:** - OpenAI's Build Hour focuses on agentic tool calling, demonstrating how to build agents that can reason and execute long-horizon tasks using the Agents SDK and new API features. - The session features live coding to build a task system, starting with a simple agent and then adding tools like 'get_user_data' and 'refund' to handle customer service scenarios. - A practical backend implementation using Flask is shown, detailing how to create SSE endpoints for streaming agent events and manage background task execution. - The concept of "tasks" is introduced as a new abstraction for long-horizon goals, broken down into agent definition, infrastructure, product interaction, and evaluation. - A key feature demonstrated is the use of "to-dos" within tasks, allowing agents to surface progress to users by updating a to-do list, enhancing transparency and user experience. - Delegation is discussed as a pattern where one agent can initiate tasks with other agents or models, enabling non-blocking operations and continuous interaction. - The session covers essential considerations for building agents, including goal specification, tool integration, infrastructure setup (runtime environments, failure handling), product interface design, and evaluation methods like LLM-based graders. **Context:** This OpenAI Build Hour session, hosted by Sarah Urbonus and Alain from the developer experience team, focuses on empowering developers with best practices and tools to scale their companies using OpenAI's APIs and models. Kicking off 2025, the series aims to provide practical, hands-on guidance, with this specific episode concentrating on the flagship models, increased context capabilities, and the "codecs" tool. The session emphasizes the practical application of these concepts through live coding and demonstrations, building upon recent OpenAI releases like the Responses API, hosted tools, and the Agents SDK. ## Detailed Analysis The Build Hour session introduces and demonstrates "agentic tool calling," a paradigm that combines an AI model's reasoning capabilities with access to tools to perform complex, long-horizon tasks. The core idea is to train models on solutions rather than specific steps, allowing them to discover optimal strategies through reinforcement learning and then use tools to act on their reasoning. This leads to goal-oriented, resourceful, and robust agents capable of executing tasks over extended periods with high consistency. The session walks through building such a system, starting with defining an agent using the Agents SDK, which simplifies the implementation of the agent loop. Alain live-codes the addition of tools, such as functions to retrieve user data or process refunds, and demonstrates how to specify end states for tasks rather than step-by-step instructions. A practical backend infrastructure is built using Flask, featuring Server-Sent Events (SSE) endpoints to stream agent progress and manage background task execution. The concept of "tasks" is elaborated as a new abstraction for long-horizon goals, encompassing agent definition, infrastructure, user interaction, and evaluation. A key demonstration involves integrating "to-dos" into the task execution flow, allowing the agent to update a user-facing progress list, making the AI's process transparent and interactive. The session also touches upon delegation, where agents can initiate tasks with other agents or models in a non-blocking manner. Finally, it covers crucial aspects like goal specification, tool management, infrastructure considerations (runtime environments, failure handling), product interface design for user engagement, and evaluation methods, including LLM-based graders. ### Tutorial - Core Concepts of Agentic Tool Calling: Reasoning combined with tools allows AI agents to perform long-horizon tasks - Models are trained on solutions, not steps, enabling emergent reasoning strategies through reinforcement learning - Agents become goal-oriented, resourceful, and robust to failures. ### Tutorial - Implementing Agents with the Agents SDK: Live coding demonstrates creating a simple agent and adding tools like 'get_user_data' and 'refund' - Specifying end-state goals for tasks is emphasized over step-by-step instructions. ### Tutorial - Backend Infrastructure for Tasks: A Flask-based backend is developed with SSE endpoints for streaming agent events - Demonstrates handling foreground and background task execution for agent processes. ### Tutorial - Task Management and Product Integration: "Tasks" are introduced as a new abstraction for long-horizon goals - "To-dos" are used to visualize agent progress in the user interface, enhancing transparency. ### Tutorial - Delegation and Background Operations: Delegation allows agents to initiate tasks with other agents or models for non-blocking execution - Discusses patterns for handing off tasks and checking in on their progress. ### Tutorial - Key Considerations for Building Agents: Covers goal specification, tool selection, infrastructure setup (runtime, failure handling), product interface design, and evaluation strategies. ### Q&A - Efficient orchestration uses Python for sequential/conditional tool calling; external memory stores (vector stores) aid long-horizon task memory management - Agent tool limits are heuristic (around 20); OpenAI hosted functions can be combined with custom functions; Responses API supports MCP for remote servers.