What OpenAI launched
OpenAI has released the Agents API in public beta, exposing the managed harness and infrastructure behind Codex to developers. The service handles the agent loop, long-running sessions, context management, tool orchestration and subagents while developers provide the task, tools and execution environment.
The practical proposition is that teams can build cloud agents without rebuilding the control plane that keeps a long-running agent alive, compacting context and coordinating parallel work.
Where the work runs
Developers can use an OpenAI-hosted sandbox, their own infrastructure, or supported sandbox providers. OpenAI lists integrations including Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel. The hosted sandbox can run code, work with files and produce artifacts.
This separation between the managed harness and the sandbox is important. It lets a team outsource orchestration while retaining control over where code executes, how files are stored and which tools the agent can reach.
Long sessions and subagents
The API includes automatic context compaction for work that spans multiple context windows, tool search that loads definitions when needed, programmatic tool calling, and multi-agent support. Subagents keep separate contexts while a coordinating agent combines their work.
Those features target exactly the infrastructure that becomes painful when an agent runs for hours rather than seconds: context growth, recovery, parallel work and the cost of carrying large tool schemas through every model call.
Early customer evidence
OpenAI’s launch material includes customer-reported results, including lower latency, lower cost per workflow and fewer failed responses in some deployments. These are useful signals, but they are supplied by launch partners and are not a common independent benchmark, so they should be treated as directional rather than universal performance claims.
What to watch
The most important question is whether the managed harness remains flexible enough for teams that need strict auditability, custom recovery logic and provider independence. Public beta also means the API surface can still move quickly. For now, the launch materially lowers the amount of agent infrastructure a developer has to own before testing a serious long-running workflow.