OpenAI Agents API Review 2026: Build Long-Running Cloud Agents with Codex and Subagents

OpenAI’s new Agents API brings the long-running agent harness behind Codex to developers through one managed API. Released in public beta on September 10, 2026, it combines durable sessions, context management, tool discovery, subagents, and configurable execution environments for applications that need to work for hours or days.

The launch addresses a difficult part of agent development: the model is only one component. Production systems also need reliable orchestration, isolated compute, state recovery, tool management, and a way to keep relevant context alive across long jobs.

What is the OpenAI Agents API?

The Agents API is a managed orchestration layer for cloud agents. A developer specifies the task, model, tools, environment, and optional multi-agent settings. OpenAI operates the Codex-based harness, while the application receives events and final output.

Key features

Durable long-running sessions

Agents can continue across extended jobs without the developer manually rebuilding every conversation. Automatic compaction preserves useful information as a session approaches its context limit, allowing work to span multiple context windows.

OpenAI-hosted or self-hosted environments

Developers can choose an OpenAI-managed sandbox, their own infrastructure, or an integrated sandbox provider. OpenAI-hosted sandboxes can run code, work with files, use installed packages, and save artifacts. Self-hosting keeps more control over networking, data location, and compute policy.

Tool search and programmatic tool calling

Instead of loading every tool definition into every prompt, tool search loads relevant definitions when needed. Programmatic tool calling can run independent calls in parallel, chain operations, filter results in code, and return only useful evidence to the model. Supported options include MCP servers, custom functions, and built-in tools such as web search.

Parallel subagents

Multi-agent support lets the main agent divide a complex assignment into independent tasks. Each subagent receives its own focused context, while the parent coordinates and combines the results. This is useful for repository analysis, research, incident response, and workflows with genuinely parallel evidence gathering.

How a session is created

A session request can name a model such as GPT-6 Astra, attach MCP tools, enable multiple subagents, identify vaults, choose an environment, and provide the initial task. The API then streams progress and output back to the application.

Pricing and availability

The Agents API is available to developers in public beta. OpenAI says there is no separate orchestration fee; customers pay for model tokens and the tools or compute they use. Beta interfaces can evolve, so production teams should pin versions, monitor costs, and build recovery paths.

Who should use it?

• Teams building coding, research, support, or operations agents that run longer than a normal request.
• Developers who need managed context compaction and durable sessions.
• Applications that coordinate several tools or parallel subagents.
• Organizations that want a choice between managed and self-hosted execution.

Limitations

A managed harness reduces orchestration work, but it does not remove the need for permissions, audit logs, evaluations, spending limits, human approval for consequential actions, and defenses against prompt injection. Multi-agent systems can also increase token use and make failures harder to trace if tasks are divided carelessly.

Verdict

The Agents API is one of OpenAI’s most consequential developer launches of 2026 because it productizes the infrastructure around the model. Teams can focus more on domain tools and workflow design, while relying on a continuously maintained Codex harness for context, tools, environments, and subagents.

Explore the official Agents API documentation

Source note: Based on OpenAI’s September 10, 2026 announcement and developer documentation. Features and pricing may change during beta.