OpenAI Opens Agents API With Managed Sandboxes
OpenAI’s public-beta Agents API lets developers run Codex-style agents with hosted sandboxes, tools, files and configurable compute.
OpenAI has put its Agents API into public beta, offering a managed way to build and run software agents that can coordinate model calls, tool use and context on OpenAI’s infrastructure.
What launched
The API includes OpenAI-hosted sandboxes where agents can execute code, manipulate files, install packages and produce artifacts. Developers can also bring their own sandbox or connect an external provider, with options for CPU, GPU and memory capacity.
The important architectural choice is that OpenAI operates the agent loop: it coordinates model calls, tool invocation and context while developers define the agent’s capabilities and execution environment. That reduces the amount of orchestration code required to move from a prototype to a persistent cloud workflow.
The launch arrives alongside a broader shift in OpenAI’s developer stack toward long-running, tool-using systems rather than single-turn text generation. Agents can now be deployed with more of the operational layer—execution, files, package installation and artifacts—already attached.
Why it matters
The API could make Codex-style agents easier to ship inside enterprise products, but it also gives OpenAI control over a valuable layer of application infrastructure. Developers gain speed and convenience; they may also become more dependent on OpenAI’s runtime, pricing and safety boundaries.
The unresolved question is whether managed sandboxes can support reliable production workloads at predictable cost. Public beta availability demonstrates a platform direction, not yet a mature operating model.