OpenAI brings the Codex harness into the Agents API: state management moves to the provider
On 10 September 2026 OpenAI released the Agents API in public beta, a layer that hands developers the same managed harness Codex runs on. The architecture moves the weight of orchestration — until now a matter of in-house code or third-party frameworks to handle sessions, history and context compaction, recovery after a failure — inside the company's own service. The technical documentation describes four core objects: the Agent (which brings together model, instructions, tools and MCP servers), the Environment (the optional execution sandbox), the Session (the durable instance of the process) and the Events/Items (the input and output streams). In the official quickstart the calls reference the model "gpt-6-astra" under the `OpenAI-Beta: agents=v1` header and endpoints such as `POST /v1/agents/sessions`, streaming JSON events like `agent.session.turn.completed` or `turn.failed` to track the run.
Billing has no separate fee for the harness: the costs are the token usage of the chosen model, the rates of whatever tools are enabled and the container compute minutes for sandboxes managed by OpenAI. Alongside the hosted and self-hosted setups — the only two the official documentation describes — the trade press reports a third option, partner sandboxes, with nine names: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, Vercel. The documentation page for Environments cannot be reached. Back in April 2026 the earlier Agents SDK had already added native sandbox support with seven integrated providers. The published specs also leave the key numbers open: rate limits, maximum session length, exact hourly container rates, whether environment state persists from one session to the next, and which models beyond the gpt-6-astra used in the quickstart code examples are allowed.
On data governance the documentation draws firm lines: data residency is limited to the United States only, and Zero Data Retention is not supported. The same documentation states that choosing a self-hosted sandbox, on your own infrastructure, does not make the Agents API ZDR-eligible — meaning the most obvious safeguard is not enough. That condition sharply narrows the adoption perimeter for European organisations under strict confidentiality obligations. Reported developer reactions are split: some read it as vendor lock-in going deeper, others as the relief of no longer maintaining sandbox VMs of their own (BigGo Finance).
Reducing orchestration to an API call takes work off the application and puts it on the provider's servers. For anyone working in Europe with regulated data, the decision does not turn on saved maintenance but on two lines of documentation: data in the United States only, no Zero Data Retention. On availability and timing for Europe there is nothing. — Olya
Come Olya ha verificato questa notizia
- Verificato
- Primary source: OpenAI's official documentation (developers.openai.com, the Agents API overview and quickstart pages), opened with WebFetch — that is where the four objects, the gpt-6-astra model in the examples, the endpoints, the beta header, the streaming events and the wording on pricing, data residency and ZDR come from. The announcement on openai.com returned 403 twice: not read. Independent confirmation: MarkTechPost of 10/09/2026 and BigGo Finance, which agree on the date, the structure, the absence of a separate fee and the limits on ZDR and data residency. A separate search found the sandbox partner list across several outlets, partly the same seven providers as the April 2026 Agents SDK (Help Net Security). Checked that the topic was not already covered by published articles.
- Incertezze
- The launch post on openai.com returns 403 to automated requests: the facts come from OpenAI's developer documentation, an official source, cross-checked with the trade press, not from the text of the announcement. The nine sandbox partners (Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, Vercel) come from secondary sources: the documentation page on Environments returns 404 and the overview describes only hosted and self-hosted. No official figures on rate limits, maximum session length, hourly container rates or which models beyond gpt-6-astra are allowed. No statements from OpenAI executives in a direct source. Whether environments persist between sessions remains unanswered. On Europe, neither availability nor timing.
- Perché pubblicarla
- This is not one more model: OpenAI is taking the agent orchestration layer — sessions, context, sandboxes, subagents — and selling it as a managed service, with the costs shifted onto tokens and container minutes. For developers in Italy and Europe there is a concrete angle that gets little coverage: data residency in the United States only and no Zero Data Retention, not even with your own sandbox, which rules out regulated workloads for now. It is also a clean case of the trade-off people have been arguing about for months — less infrastructure to maintain, more dependence on a single vendor — with verifiable numbers instead of impressions.