OpenAI Launches Agents API and Hosted Sandboxes in Public Beta

OpenAI has launched the Agents API in public beta for all developers. [1] The service manages the agent loop on its infrastructure by coordinating model calls, tool use, and context management. [1] Developers pay for the tokens and tools their agents use, with no additional fees for the Agents API itself. [1] The system handles long-running sessions while allowing developers to control the agent's capabilities. [1]

OpenAI introduced hosted sandboxes that allow agents to run code, work with files, and produce artifacts. [1] These OpenAI-hosted sandboxes use standard container rates and allow users to install packages or add skills. [1] Alternatively, users can bring their own sandbox or connect to providers like Modal, Vercel, and Oracle Cloud. [1] Self-hosted options support fully managed environments or deployments within a virtual private cloud (VPC). [1]

By absorbing the orchestration and context management layers into a native API, OpenAI is commoditizing the middleware that developers previously had to build from scratch. The explicit support for first-class integrations with providers like Modal and Vercel positions OpenAI's infrastructure as a central hub for agentic workflows rather than a strict walled garden. However, the separation of model usage and standard container billing could lead to unexpected expenses for long-running agents, requiring developers to carefully calculate compute costs upfront. What the sources don't address: Whether the Agents API includes built-in guardrails to prevent runaway compute loops inside the new hosted sandboxes.