OpenAI has opened an API for cloud agents with durable sessions, tools and sandboxes. Understand what to outsource and what remains under company governance.

Direct answer

The Agents API, released in public beta by OpenAI on September 10, 2026, offers the Codex harness as a managed service to run long-running agents with tools, context, and workspaces. It reduces the infrastructure that each company needs to build, but does not transfer responsibility for scope, data, permissions, evaluations, observability and approval of actions.

What the managed tier offers

The API brings together persistent sessions, context management, tools, files and environments where the agent can work. The proposal is to prevent each team from rebuilding all the necessary orchestration for tasks that last hours or days.

Environment is a risk decision

The company can consider hosted sandbox, its own infrastructure or compatible partners. The choice should reflect data sensitivity, dependencies, network requirements, region, cost, and ability to investigate after an incident.

Tools expand power and attack surface

MCP, functions, and search allow the agent to query and change systems. Use specific credentials, least privilege, allowlists, spending limits, and confirmations before external or irreversible actions.

Assessments need to represent actual work

Test for normal cases, ambiguities, tool failures, resume after interruption, and branch attempts. Measure validated completion, time, cost, human interventions, incidents, and ability to produce reproducible evidence.

Nexus Reading

Managed infrastructure accelerates the first agent, but differentiation comes from proprietary context, well-designed tools, and governance. The pilot must prove value without transforming technical autonomy into unrestricted decision-making autonomy.

FAQ

Does the Agents API replace the entire agent infrastructure?

It manages an important part of the harness and sessions, but the application still defines data, tools, permissions and controls.

Is it possible to use your own infrastructure?

The announcement describes options for an environment hosted by OpenAI, either its own or integrated with partners.

How to start safely?

Choose a verifiable task, limit access, and require human approval for material actions.

Essential guides to delve deeper into the decision

Primary sources and references

This editorial analysis was produced by Nexus from the official sources below, consulted on September 11, 2026. The text is original and interprets practical implications for companies.

Date reported by the main source: September 10, 2026.