Build Hour: Agents SDK

### TL;DR OpenAI released a major update to the Agents SDK, introducing a Codex-style runtime framework, separation of sandbox and runtime framework, memory, skills, and more. It also launched Hosted Shell tools, Container Endpoints, Skills API, and a TypeScript version, aiming to simplify the construction and deployment of production-level agents. --- ## Agents SDK: From Zero to Production-Grade Evolution In the latest Build Hour session, OpenAI’s Christine (Startup Marketing team) and Steve (API team engineer) detailed the core updates to Agents SDK over the past few months. Steve emphasized that the autonomous working capabilities of models are rapidly improving—whether it's Codex coding agents, security scanning agents, or internal data agents, they can remain efficient in long-running tasks. However, deploying agents in production still faces challenges: How to maximize performance while maintaining cross-model flexibility? How to isolate execution environments and manage state? How to achieve a highly customizable framework? The Agents SDK updates are designed to address these pain points. ### Codex-style Runtime Framework: Making Agents Work Like Codex The Agents SDK introduces a "Codex-style" runtime framework, generalizing Codex's advantages in coding scenarios (such as automatic compression, async Shell interaction, task tracking) to any domain. Core elements include: - **Agent Loop**: Automatically handles LLM calls, tool execution, and context updates; developers do not need to build the orchestration layer themselves. - **Web search, file search, MCP, code interpreter, skills** are available out of the box or can be enabled with one click. - **Computer use**: Models can write Shell commands and use an asynchronous interaction loop—they can wait when a command is not yet complete, then return to process, and track running commands in real-time. ### Separation of Sandbox and Runtime Framework: Key to Production Deployment In the past, Codex’s sandbox (e.g., local notebooks) was tied to the runtime framework, causing issues like container death, state loss, and key exposure in production environments. Agents SDK decouples the two: - **Sandbox** becomes a fully ephemeral isolated environment that can expire or be rebuilt at any time, with no need to worry about state persistence. - **Runtime framework** can be deployed on existing infrastructure like Temporal, AWS, etc., handling reloading, snapshots, and multi-tenant session management. - Provides first-class sandbox support: platforms like ETB, Modal, Cloudflare, Vercel, Docker can be integrated. ### New Features: Skills, Memory, Containers - **Skills API**: By uploading skill packs (containing skill MD files and scripts), create a central trusted source. For example, a tax preparation skill defines all IRS rules and document processing scripts. Supports version management, setting default versions, and seamless collaboration with Hosted Shell. - **Agent memory**: Tasks can improve over time, supporting cross-session context retention. - **Containers**: A new concept in the API and SDK for running file operations in isolated environments. --- ## New API Releases: From Lightweight to Production ### Hosted Shell Tool (in Responses API) A "lightweight sandbox" model: with just one API call, upload files, and the model can write code, process files, and return results in a temporary container. Supports automatic mode (auto start/stop container) or manual use of Container Endpoints (create, upload, attach to a response request). ### Container Endpoints and Network Control - Create container, upload files, start container, then attach to the Responses API. - Supports domain whitelisting or fully locking down the container (no inbound/outbound) for enhanced security control. ### TypeScript Version Release Following the Python version in April, the TypeScript version of the sandbox agent functionality is now available, meeting the needs of TypeScript developers. --- ## Live Demo: Agent Task Tracker In the main segment of the session, Steve built an agent task tracker based on Agents SDK live. The demo showed how to quickly implement task assignment, status tracking, file operations, and other functions using the new framework. Since the demo took up most of the livestream, details can be found in the recording. --- ## Q&A and Summary At the end of the livestream, ample time was arranged for Q&A, with special guests answering questions about multi-tenant implementation, memory mechanisms, MCP integration, container resource limits, and more. Christine mentioned that recordings of all past sessions can be found on the OpenAI website, and a replay of this livestream will also be available. --- **Source**: Build Hour: Agents SDK - YouTube (https://www.youtube.com/watch?v=tK32trvj_b4)