OpenAI Releases Agents API in Public Beta with Hosted Sandboxes

OpenAI released the Agents API in a public beta for all developers. [2] The system puts the Codex harness behind a single API call and manages orchestration, long-running sessions, and context. [1][2] OpenAI handles the agent loop on its own infrastructure to coordinate model calls and tool usage. [2] Developers do not pay additional fees for the Agents API itself, but are charged for the tokens and tools their agents utilize. [2]

The release includes the introduction of OpenAI-hosted sandboxes, which allow agents to run code, work with files, and produce artifacts. [2] Users can supply files, install packages, and add plugins while OpenAI provisions and manages the environment. [2] Alternatively, developers can utilize self-hosted environments or connect a sandbox provider, with options for CPU, GPU, and memory configurations. [2] The system offers integrations with providers including DigitalOcean, Oracle Cloud, Vercel, Cloudflare Dev, and Modal. [2]

The new Agents API abstracts away the complexity of agent orchestration by managing the agent loop and context windows natively. [2] By offering both fully managed OpenAI sandboxes and integrations with third-party providers like DigitalOcean and Oracle Cloud, the platform caters to diverse infrastructure requirements. [2] What the sources don't address: How the performance and cost of the managed Agents API compare to self-orchestrated solutions built on open-source frameworks.