The OpenAI Agents API public beta opened on 10 September 2026, and it gives every developer the managed Codex harness as a plain API: OpenAI runs the agent loop on its own infrastructure — coordinating model calls, tool use and context — whilst you supply the tools, pick the execution environment and pay only for the tokens and tools your agents consume, with no additional fee for the API itself. That comes straight from OpenAI's official announcement, "Introducing the Agents API", published 10 September 2026, and the same-day entry in the official OpenAI API changelog.
📺 Watch: I Automated SEO With GPT-6 Astra + Codex
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside · Want AI SEO help 1-on-1? Book a free SEO strategy session →
For anyone building AI automation, this is a significant shift. Until now, if you wanted a long-running cloud agent you either wired together your own harness — session handling, context management, retries, sandboxing — or you adopted an open-source agent framework and hosted it yourself. The Agents API moves that whole layer onto OpenAI's side of the fence. According to the announcement, the pitch is that useful agents need "a powerful harness that manages context, uses tools efficiently, and coordinates subagents", plus infrastructure that keeps them running reliably for days — and OpenAI now sells exactly that as a managed service.
This article walks through what shipped, what it costs, how the sandbox options work, and how it compares with running your own agent stack — every claim sourced to OpenAI's own announcement, changelog and docs, with links so you can verify each one.
How The OpenAI Agents API Public Beta Works
The core of the release is the Codex harness — the same machinery that powers OpenAI's Codex coding agent — exposed behind an API call. Per the official announcement, developers can "build and run cloud agents with the Codex harness, fully managed by OpenAI", with OpenAI handling "orchestration, long-running sessions, and context management".
The official API changelog entry for 10 September 2026 breaks the Agents API public beta into four concrete capabilities:
- A managed harness — OpenAI handles session orchestration, context compaction and recovery, so you do not write that plumbing yourself.
- Durable sessions — agents continue work across turns, with progress streaming back to your application as they run.
- Custom tools and MCP servers — you connect your own tools and Model Context Protocol servers, so an agent can reach your data and systems.
- Flexible sandboxes — agents run in OpenAI-hosted sandboxes, or in sandboxes on your own infrastructure or from supported partner providers.
Context compaction matters more than it sounds. Long-running agents die by context: the session fills up, quality degrades, and most home-grown harnesses handle it badly or not at all. Having compaction and recovery managed by the provider is a genuine chunk of engineering you no longer own. If you want to understand what a harness actually does all day — and why harness quality changes agent results as much as model quality does — the Goldie Bench write-up covers how different agent brains and harnesses compare on real tasks.
If you want to turn agent tooling like this into an actual income stream rather than a weekend experiment, that is exactly what we build inside the AI Profit Boardroom — daily tutorials, prompt libraries and weekly live coaching calls: see how members are shipping agent systems. And if you want a plan for your own site first, book a free SEO strategy session and get it mapped out 1-on-1.
Sandboxes: OpenAI-Hosted, Self-Hosted, Or A Partner
Every agent needs somewhere to actually do the work — run code, edit files, keep intermediate results. The Agents API gives you three routes, per the announcement:
| Option | What it is | Who runs the compute |
|---|---|---|
| OpenAI-hosted sandbox | Managed environments where agents can run code, work with files and produce artifacts; you can supply files, install packages and add skills and plugins | OpenAI |
| Self-hosted sandbox | Bring your own infrastructure and connect it to the harness | You |
| Partner sandbox | Integrated third-party sandbox providers | The partner |
The announcement names nine integrated sandbox partners at launch: Blaxel AI, Cloudflare Dev, Daytona, DigitalOcean, E2B, Modal, Oracle Cloud, Runloop AI and Vercel. That list is worth pausing on — it means the harness layer and the compute layer are now decoupled, so teams already running workloads on, say, Cloudflare or DigitalOcean can keep their compute where it is and still hand the agent loop to OpenAI.
What The Agents API Costs
Pricing is unusually simple for a beta of this size. Per OpenAI's announcement: there are no additional fees for using the Agents API — you pay for the tokens and tools your agents use. OpenAI-hosted sandboxes are the one extra line item: they are billed at standard container rates, separately from model API charges. Self-hosted sandboxes keep compute costs on your own bill, and partner sandboxes are priced by the partner.
In practice that means the cost model of an Agents API workload looks like any other heavy API workload — dominated by token spend — plus container time if you use OpenAI's environments. If you are already tracking per-task token costs for your agents (and you should be), nothing about this billing structure will surprise you. The same discipline applies here as anywhere else in the stack: measure cost per completed task, not cost per call.
Agents API Vs Running Your Own Harness
The obvious question for anyone in this space: should you use OpenAI's managed harness, or keep running an open agent framework on your own machine or VPS? The honest answer is that they solve different problems, and the OpenAI Agents API public beta does not make self-hosted agents obsolete.
Where the managed harness is strong
Reliability engineering is the headline. Long-running sessions, recovery after failures, context compaction and progress streaming are exactly the parts of agent infrastructure that are tedious to build and brutal to debug. If your agents are OpenAI-model-based anyway and your workload is bursty cloud work — research runs, coding tasks, batch jobs — the managed route removes a whole category of operational pain, and the sandbox partner list means you are not locked into OpenAI's compute.
Where self-hosted still wins
Control and model choice. A self-hosted agent framework lets you mix providers, run local models, own your data path end to end, and customise the loop itself. It also keeps you portable: your workflows are not expressed in one vendor's API. If you are weighing up the self-hosted route, start with the guide to the best free autonomous AI agents, and if you run Hermes, the Connect Codex to Hermes guide shows how the two worlds already talk to each other. The structured way we organise agent workflows on the self-hosted side is Agent OS — the operating layer that turns a pile of agent experiments into a repeatable system.
MCP is the bridge
Notably, the Agents API supports connecting MCP servers — the same Model Context Protocol used across Claude, Hermes and most serious agent stacks. Tools you have already wrapped as MCP servers are not stranded on either side of this decision; the Hermes MCP server guide covers how that pattern works in practice. Building your tool layer on MCP is increasingly the safest bet, because both the managed and self-hosted worlds can consume it.
📺 Watch: I Found a FREE Way to Run Codex
How To Get Started With The Agents API
Access is open: the beta is available to all developers, per the announcement. OpenAI points newcomers at the overview documentation and the quickstart guide on its developer portal. A sensible first project is something you already understand well enough to judge — a research task, a repo chore, a content pipeline step — run first in an OpenAI-hosted sandbox (least setup), then moved to self-hosted compute if the economics or data requirements demand it.
One related note on the OpenAI side of your stack: if you are experimenting with Codex itself, there are legitimate free routes worth knowing about before you spend — the free OpenAI API key for Codex guide covers the current options, and the Codex course takes you from zero to working agent workflows. For the pattern of pairing Codex with a Hermes agent as the orchestrator, see Hermes agent + Codex.
What This Means For AI Automation Builders
The Agents API is OpenAI formalising something the automation community has known for a year: the harness is the product. Models are increasingly interchangeable; the layer that manages context, tools, sessions and recovery is where reliability — and therefore business value — actually lives. OpenAI charging nothing extra for the harness and monetising tokens and container time tells you where they think the volume is going: long-running cloud agents doing real work, not chat.
For builders, the practical takeaway is to design workflows that are harness-portable. Keep tools in MCP servers, keep prompts and procedures in files you own, and treat the harness — OpenAI's, Hermes, or anything else — as swappable infrastructure. Do that and every new release like this one is an option you can exercise, not a migration you have to survive.
If you want working agent systems — not just news about them — the AI Profit Boardroom includes the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls: get inside the AI Profit Boardroom. Want AI SEO help 1-on-1 instead? Book a free SEO strategy session and we will map your next 90 days.
