OpenAI Opens Managed Codex Harness With New Agents API

OpenAI has opened the Agents API in public beta, giving developers an API-level way to build and run cloud agents around the same managed Codex harness used by OpenAI’s own agent products. The announcement is a notable shift from shipping isolated model calls toward offering the operational layer around long-running agent work.

What OpenAI released

The Agents API manages sessions, orchestration, context compaction and recovery while the application supplies instructions, tools and the execution environment. OpenAI’s developer documentation describes agents that can run commands in a sandbox, edit files, connect to MCP servers and produce artifacts.

The API supports OpenAI-hosted sandboxes as well as self-hosted environments. The managed harness can also steer an agent while it works, resume a session, and delegate subtasks to subagents. Model usage remains billed at the selected model’s API rates, while sandbox and tool usage follow their respective pricing.

Why it matters for builders

The practical change is less about a new model and more about removing infrastructure that every serious agent team eventually has to recreate. Session persistence, context management, failure recovery and tool orchestration are difficult to make reliable once an agent runs longer than a single request.

For automation teams, the API creates a clearer boundary: the platform can own the agent runtime, while the application owns business logic, permissions and integrations. That is useful for workflows that need durable state or human intervention rather than a one-shot response.

The trade-off is control. A managed harness can accelerate delivery, but production teams still need explicit egress policies, credential isolation, audit logs and approval gates around consequential tools. MCP connectivity and sandbox access expand what an agent can do, so they also expand the blast radius of a misconfigured session.

OpenAI’s September 10 announcement positions the Agents API as a shorter route from an idea to a working cloud agent. The next test is whether its managed defaults can satisfy teams that need portability, predictable costs and evidence that every tool action is accountable.