OpenAI released the Agents API in public beta on September 10, 2026, at 8am Pacific, the same morning as GPT-Live-1 and ChatGPT for Financial Services. The subtitle is the whole announcement: “Build and run cloud agents with the Codex harness, fully managed by OpenAI.” The post explains the reasoning in one sentence we agree with word for word: “Useful agents need a powerful harness that manages context, uses tools efficiently, and coordinates subagents.”
This page is the record: what shipped, how the environment choice works, what it costs, and what it means for a company that builds its own harness. Every quotation is from OpenAI’s launch post or the Agents API overview in the developer docs, both read the day of release.
On this page · 8 sectionsOpen
OpenAI released the Agents API in public beta on September 10, 2026. Its one-line pitch: ‘Build and run cloud agents with the Codex harness, fully managed by OpenAI.’ The harness that runs Codex is now something you call, not something you run.
You choose where the agent computes. OpenAI’s words: ‘OpenAI hosts and maintains the harness. You choose the agent’s compute environment: in an OpenAI-managed sandbox, on your own infrastructure, or with one of our sandbox partners.’ Nine partners are named, from Cloudflare to Vercel.
The harness is a moving target you no longer chase. ‘The Agents API provides versioned access to these capabilities with each model launch’, so a new model arrives with the harness already tuned for it.
Long work is handled for you. The API ‘automatically compacts earlier context as a session approaches its context limit’, tool search loads tool definitions only when needed, and programmatic tool calling ‘lets agents run calls in parallel, chain related operations, and filter or combine results in code’.
Subagents are built in. The main agent can ‘break complex tasks into independent pieces and delegate them to subagents that work in parallel’, and ‘Each subagent maintains its own context’.
It costs nothing on top. ‘There are no additional fees for using the Agents API’; you pay model tokens at the model’s API rates, tools at their standard rates, and OpenAI-hosted sandboxes at container rates.
It is open at the bottom and managed at the top: ‘powered by the open-source Codex harness’, while ‘OpenAI operates and maintains that harness’. Public beta today for all developers; general availability is the stated destination.
§ 01What shipped
| Item | Detail |
|---|---|
| What it is | “The Agents API gives your application access to the Codex harness through an OpenAI-managed API.” |
| Division of labor | “OpenAI manages sessions, orchestration, context compaction, and recovery while your application provides tools and chooses its execution environment.” |
| One call | client.beta.agents.sessions.create with an agent (model, instructions, tools, MCP servers), an environment, and the input |
| Environments | OpenAI-hosted sandbox, self-hosted, or a partner sandbox |
| Tools | MCP servers, custom functions, built-in tools such as web search; tool search and programmatic tool calling |
| Multi-agent | Subagents with their own context, coordinated by the main agent; the samples cap concurrency at three or four |
| Status | Public beta for all developers; “we work toward general availability” |
| Price | No API fee; model tokens, tools and OpenAI-hosted container time at standard rates |
§ 02The harness, as OpenAI now describes it
The docs list what the managed harness does. Set beside the launch post’s sentences, it is the plainest description OpenAI has published of what a harness is for.
| Job | OpenAI’s line |
|---|---|
| Run commands and code | Agents “can execute code, edit files, connect to MCP servers, and produce artifacts” in a sandbox |
| Apply skills and instructions | Capability directories loaded from the environment, for example /workspace/capabilities/skills |
| Reach external data | “The Agents API supports MCP, custom functions, and built-in tools like web search.” |
| Steer mid-flight | “Send another task to the same session, or guide the agent during its current turn.” |
| Manage context | “automatically compacts earlier context as a session approaches its context limit, preserving information the agent needs to continue” |
| Delegate | “break complex tasks into independent pieces and delegate them to subagents that work in parallel” |
| Resume | “Resuming a session where it left off.” |
Two of these deserve a closer look because they are the parts every harness builder has been writing by hand. On tools: tool search “loads relevant tool definitions as needed, helping reduce token usage and cost while preserving the model’s cache”, and programmatic tool calling “lets agents run calls in parallel, chain related operations, and filter or combine results in code”. On subagents: “Each subagent maintains its own context, helping it stay focused on its assignment, while the main agent coordinates their work and brings the results together.”
§ 03Your sandbox or theirs
The environment is the one thing OpenAI hands back to you. “Different workloads need different compute, storage, and deployment options. The Agents API lets you choose a sandbox that fits your application.” Three paths:
| Environment | What OpenAI says | Billing |
|---|---|---|
| OpenAI hosted sandbox | “leverages the same sandboxing infrastructure that powers Codex and ChatGPT”; configured with your files, packages, skills and plugins | Standard container rates |
| Your own infrastructure | Self-hosted; your application chooses the execution environment and provides the tools | Yours |
| Partner sandboxes | Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, Vercel: managed or in-VPC, specific storage, “Different CPU, GPU, and memory configurations” | The partner’s |
The pitch to platform teams is that the harness stops being a maintenance job: “Taking advantage of new model capabilities often means reworking your harness, taking valuable time away from improving your application.” Versioned access “with each model launch” is the answer, and it is a real one. Anyone who has rewritten a tool loop the week a new model shipped knows exactly what is being sold.
§ 04What it costs
The API itself is free. “There are no additional fees for using the Agents API”; instead “you simply pay for the tokens and tools your agents use”. The docs are precise about the three lines on the bill: “Model usage is billed at the selected model’s API rates. OpenAI tools use their standard rates, and OpenAI-hosted sandboxes use standard container rates.” A self-hosted or partner environment moves the third line to whoever runs the box.
That shape matters more than any number. The harness, the part that took teams the longest to build, is priced at zero, and the meter runs on the model. It is the same bet as GPT-Live-1’s per-second voice clock the same morning: sell the machinery cheaply, meter the intelligence.
§ 05The week the harness moved
The Agents API landed in a seven-day stretch where OpenAI’s agent stack changed more than its model lineup did.
| Date | Event |
|---|---|
| September 3 | GPT-6 Astra launched; Codex CLI 0.153.0 released |
| September 9 | Codex CLI 0.154.0 released, Astra in the model picker |
| September 10 | Agents API public beta; GPT-Live-1 in the API; ChatGPT for Financial Services |
§ 06Where the employee lives
Our conflict, declared: we build CellCog, an AI employee platform, and it runs on a harness of our own, built since October 2025 and described in our engineering post. Ours is not open source; we publish what it does and why, never how. Our agents at the Core and Max tiers run Claude Fable 5.1, with Gemini 3.8 Flash at the Flash tier, and nothing routes through OpenAI’s harness. So read what follows as a competitor’s view of a very good launch.
The Agents API confirms the thing we have been saying since we started: the harness is the product. OpenAI now names it, prices it at zero, and runs it for you. What it sells is a session: a task, a model, some tools, a box to compute in, and a harness that keeps the session alive for hours. That is the right unit for a developer building an agent into their own product, and it is a different unit from the one a business owner needs.
| Agents API session | CellCog AI employee | |
|---|---|---|
| Who it is for | A developer building an agent into their product | A business owner hiring for a role |
| The unit | A session working a task, resumable | An employee with a role, whose working sessions carry forward |
| Harness | The open-source Codex harness, operated by OpenAI | Our own harness, operated by us; the why and what are public, the how is not |
| Memory | Compaction inside a session; sessions can be resumed | Written records that carry between working sessions: notes, a task board, an inbox, a handover |
| Working together | Subagents spawned inside a session, each with its own context | Coworkers with names and roles who message each other, form teams, and report to a person |
| Environment | An OpenAI sandbox, your infrastructure, or a partner | Each employee’s own computer and browser on our servers, with separate logins per employee, plus the tools you connect |
| Guardrails | Your application provides the tools and decides what they may do | Every command that reaches your world is classified by the agent; the platform rejects any command that arrives unclassified, and anything above your threshold waits for your approval |
| Pricing | Tokens, tools, container time | Pay for the work, not the hire; the cost depends purely on how much work you assign (pricing) |
None of this is a criticism of the API. A session with a great harness is exactly what a developer should be able to buy in one call, and until yesterday they could not. It is a statement about what sits on top: the employee, the part with a name, a memory and a manager, is what we spent the last eleven months building, and OpenAI’s launch post is the clearest outside evidence yet that the layer underneath it was worth building well.
§ 07What we are watching for
- General availability. The post ends with “During the public beta, we’ll iterate quickly based on your feedback as we work toward general availability.” Any GA notice, rate limit or session-length limit published in the docs gets folded here.
- The supported model list. Today the samples say gpt-6-astra and the post promises versioned access per launch; the first non-OpenAI or non-Astra model in the docs is a record change.
- Pricing rows. A container-rate table specific to Agents API sandboxes, or any fee for the API itself, flips the cost section.
- Codex CLI convergence. Whether the open-source harness and the managed one stay in step release by release, starting with Codex 0.155.
§ 08Sources
OpenAI, Introducing the Agents API, September 10, 2026, 8am Pacific (every quotation above is from this post or the Agents API overview and pricing pages in OpenAI’s developer docs, all read September 11, 2026). Codex CLI release dates from the GitHub releases page. CellCog routing statements reflect our configuration on the same date.
Q1Which models does it support?
OpenAI’s launch post names none as a list; its code samples use gpt-6-astra. The post says the API ‘provides versioned access to these capabilities with each model launch’, so treat the supported set as whatever OpenAI’s docs list on the day you build.
Q2What are the four concepts?
Per the docs: an Agent (model, instructions, tools and MCP servers), an Environment (an optional sandbox or computer), a Session (a durable instance of an agent working on tasks) and Events and items (the inputs sent and the output produced). A session is created, given a task, streamed or webhooked for progress, then continued or steered.
Q3Who are the sandbox partners?
Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel, per the launch post, covering managed environments, deployments inside your VPC, specific storage mechanisms, and different CPU, GPU and memory profiles.
Q4What did the first customers report?
OpenAI quotes eight design partners. Ciridae’s CTO says ‘our evaluation score went from 0.71 to 0.85’ and that ‘the new APIs gave us a 4x latency reduction’ on subagent flows. These are the vendor’s chosen quotes; no independent run existed on launch day.
Q5Does CellCog use the Agents API?
No. CellCog runs its own harness, and our agents at the Core and Max tiers run Claude Fable 5.1, with Gemini 3.8 Flash at the Flash tier; none of it routes through OpenAI’s harness. We publish what our harness does and why, never how it is built.