The OpenAI Agents API is now open to every developer, no waitlist required, as of September 10. It hands you the same engine that runs Codex and ChatGPT’s background tasks — sandboxed code execution, automatic memory management, and the ability to split one job across a team of AI subagents — through a single API call.
If you’ve ever tried to build an AI agent that keeps working for hours without falling over, you know how much plumbing that takes. OpenAI just packaged that plumbing and started selling it by the token.
Here’s the 10-second version before we get into how it actually works:
- Launched September 10, 2026 in public beta — no invite or waitlist needed
- OpenAI’s Codex harness now handles context management, tool loading, and multi-agent coordination for you instead of you writing that logic yourself
- No separate API fee — you pay for model tokens, tools, and sandbox time (the example model, GPT-6 Astra, costs $10 per million input tokens and $50 per million output tokens)
- Runs on OpenAI’s hosted sandbox by default, your own servers, or partner clouds like Cloudflare, DigitalOcean, E2B, and Vercel
- US-only data residency for now, and Zero Data Retention isn’t supported yet
What Is the OpenAI Agents API, Actually?
Think of it like the difference between buying a car and buying an engine. OpenAI’s ChatGPT app and Codex already run autonomous agents that write code, browse files, and work across multiple steps without you watching every move. The Agents API is OpenAI taking that same engine out of the car and selling it to you directly.
Before this, if you wanted an agent that could run for an hour, remember what it did ten steps ago, and hand a subtask to a helper, you had to build all of that yourself. Now it’s one API call, and OpenAI’s servers do the coordinating. The example in OpenAI’s own docs runs on GPT-6 Astra, though the harness isn’t limited to a single model.
The Codex Harness Does the Boring Work for You
The part doing the heavy lifting is what OpenAI calls the “harness” — basically an operating system for your agent. Two things make the biggest practical difference.
First, context compaction. As a long-running agent session fills up with messages, tool results, and file contents, it eventually hits a token limit. The harness automatically summarizes older parts of the conversation so the agent keeps working instead of crashing or losing track of why it started.
Second, lazy tool loading. Instead of stuffing every possible tool definition into the prompt from message one, the harness only loads a tool’s instructions when the agent actually needs it. That alone can meaningfully cut your token bill on agents with a large toolset.
Sandboxes: Where Your Agent Actually Runs
Every agent needs somewhere to execute code and touch files, and OpenAI gives you three options instead of locking you into one.
The default is OpenAI’s own hosted sandbox — a contained environment where the agent can run code and produce files without touching your infrastructure at all. Teams that need more control over network access, credentials, or data location can run the harness against their own servers using codex exec-server, which only opens outbound connections. And if you’d rather not manage either, OpenAI has built-in integrations with Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel.
That flexibility matters more than it sounds. A startup shipping a quick prototype can use the hosted option and move on the same day. A company with compliance requirements can keep everything inside its own network instead of trusting a black box.
Subagents: One Request, a Team of Workers
The feature developers seem most excited about is subagent delegation. Turn it on, and your main agent can break a big task into independent pieces and hand each one to its own subagent, each with separate context.
OpenAI’s own example caps this at three subagents running at once, though that’s a setting you control. The obvious use case is incident response: one subagent checks recent deployments, another digs through error logs, a third looks at dependency changes — all at the same time, instead of one agent working through the list in order.
Early adopters are reporting real numbers. Fraud-review platform SafetyKit says it cut its cost per case by 60% after switching over. Ciridae reported a 4x latency drop on subagent workflows, and logistics company Nash.ai is now running thousands of agents at once across its network.
Agents API vs. Agents SDK vs. Responses API — Which One Do You Need?
OpenAI now sells three different ways to build with agents, and it’s easy to grab the wrong one. Here’s how they actually compare.
| Tool | Best For | Who Manages Infrastructure | Setup Effort |
|---|---|---|---|
| Agents API | Long-running, production agents | OpenAI (hosted harness) | Low — one API call |
| Agents SDK | Custom orchestration logic | You, in your own code | Medium — you write the loop |
| Responses API | Single-turn tool calls | You | Low, but no persistent agent memory |
If you’re not sure which row fits you, you probably want the Agents API. It’s the one built for agents that need to keep running, remembering, and delegating without you babysitting the process.
Frequently Asked Questions
What is OpenAI’s Agents API used for?
It’s for building AI agents that run autonomously for extended periods — writing code, researching, or automating multi-step workflows — without you managing memory, tool loading, or task delegation by hand.
How much does the Agents API cost?
There’s no separate platform fee. You pay for the model tokens, any billable tools, and sandbox or container time you actually use, based on standard OpenAI API pricing.
Is the Agents API different from the Agents SDK?
Yes. The Agents SDK gives you code libraries to build your own orchestration logic, while the Agents API is a managed service where OpenAI runs that orchestration for you.
Can I run OpenAI agents on my own servers?
Yes. You can attach the harness to your own infrastructure using codex exec-server, or use one of OpenAI’s partner sandbox integrations instead of the default hosted option.
Is the Agents API available outside the US?
You can call it from anywhere, but data residency is currently US-only, and Zero Data Retention isn’t supported yet — both worth checking if you’re under strict compliance rules.
The Agents API isn’t going to change what you see in ChatGPT tomorrow, so if you’re not writing code for a living, feel free to skip this one. But if you are, it’s a genuinely useful shortcut: the pricing model means you’re not paying a platform tax on top of tokens, and skipping the orchestration work that used to eat a sprint or two is a real time save. It’s the same bet Microsoft made with its unified Copilot agent experience — hide the plumbing, sell the outcome — and for once, the beta label seems to mean “still improving,” not “barely works.”
Leave a Reply