OpenAI introduced the Agents API in public beta on September 10, 2026. It gives developers access to the managed Codex harness: the software around a model that coordinates its work. OpenAI runs that harness, while developers choose tools and an execution environment. OpenAI’s announcement
This matters for teams whose agent prototype has accumulated infrastructure of its own. Once a task needs to survive a disconnect, return to earlier work, or produce files across several steps, there is considerably more to operate than a prompt. The decision is whether maintaining that machinery is a useful part of your product.
What the service takes over
OpenAI’s overview describes managed sessions, orchestration, context compaction, and recovery. A session persists across turns; applications can send more input, follow events, and steer ongoing work. The agent’s configuration defines its model, instructions, and tools, while an optional environment supplies a place to run commands and work with files. Agents API overview
The launch also includes tool search, programmatic tool calling, and subagents. Tool search loads relevant definitions as needed. Programmatic calling lets the agent combine operations in code. Subagents can work on independent assignments with their own context. OpenAI says the underlying Codex harness is open source. Harness capabilities
Those capabilities are useful candidates for a document-review workflow: separate reviewers could examine different sections before a coordinating agent assembles the result. That is a proposed application, not a measured speedup. Parallel work still needs a final check for conflicting conclusions and missing evidence.
Choosing where code runs
The API supports OpenAI-managed environments, your own infrastructure, and sandbox partners. OpenAI lists providers including Cloudflare, Daytona, E2B, Modal, and Vercel. Environment choices
An OpenAI-hosted sandbox supplies a Linux workspace with Python, Node.js, and command-line tools. Configuration can add packages, files, setup commands, skills, and plugins. The documentation points developers toward self-hosting when they need their own image, compute, or private network. Hosted sandbox documentation
A durable session should not be treated as permanent file storage. The hosted sandbox can be deleted after an hour without activity or keep-alives, and closing the event stream does not cancel its task. An application should retrieve outputs it needs to retain and make cancellation an explicit operation. Sandbox lifecycle limits
Costs and data boundaries need an early check
OpenAI says there is no extra Agents API fee. Usage still carries charges: the overview specifies model rates, tool rates, and container rates for OpenAI-hosted sandboxes. A managed harness therefore changes what your team operates without eliminating execution costs. Launch pricing · Billing details
The same overview currently limits Agents API data residency to the United States and says it does not support Zero Data Retention. A self-hosted sandbox does not change that eligibility. These are constraints to check before sending a real workload, particularly if choosing your own compute is part of a data-control requirement. Data controls
A useful first evaluation
The official quickstart builds an agent that writes and runs a directory-tree script. It requires an application API key with agent read/write permissions and Responses write permission, and instructs developers to keep the key outside the sandbox. Agents API quickstart
For a product evaluation, choose one bounded task with a result a teammate can verify: a report from a supplied file, for example. Test continuation after a disconnect, missing inputs, output retrieval, and the point where the agent must request a decision. Record the entire task’s cost, including failed attempts. Our recommendation is to compare those results with your existing implementation before moving a broader workflow.
The GPT-Live-1 API release addresses how users speak to an application. The Agents API addresses how an agent’s work runs. Together they invite a conversational interface over longer tasks, but each part needs its own evidence of success: a correct response to the user and a verified result from the work.