The model was never the hard part. Anyone who has tried to ship a long-running AI agent knows the real work lives in the unglamorous plumbing underneath: resuming a session after a crash, compacting context before the window fills, retrying a timed-out tool call, keeping a fleet of subagents from trampling each other. On September 10, OpenAI started renting that plumbing out.
The company opened the public beta of its Agents API, handing every developer the same managed harness that runs Codex and ChatGPT for Work. OpenAI keeps the session alive, compacts the context as it approaches the limit, and recovers after failures. The developer supplies instructions, tools, and MCP servers, then picks where the code actually executes. Every request carries the beta header agents=v1, and the documentation samples run on gpt-6-astra. “As we’ve scaled Codex and ChatGPT for Work to millions of people around the world,” the company wrote, “we’ve learned what it takes to make long-running agents work well in practice.”
What actually shipped
The surface area is deliberately small. Four objects do the work: an Agent (model, instructions, tools, MCP servers), an optional Environment sandbox, a durable Session, and the stream of events and items that session emits. Inside a session an agent can run code, edit files, search the web, load skills from a capability directory, produce artifacts, and fan work out to subagents under a concurrency cap. OpenAI's own sample incident-response agent sets max_concurrent_subagents to 3 and tells the agent to write its findings into /workspace/outputs.
Two harness features OpenAI has been grinding on internally come along for the ride. Tool search loads tool definitions only when the model needs them, which trims token spend while preserving the prompt cache. Programmatic tool calling lets an agent run calls in parallel, chain them, and filter or combine results in code, so a query over a large dataset returns only the relevant rows into context rather than the whole payload.
Environment choice is the real architectural decision. An OpenAI-hosted sandbox uses the same infrastructure behind Codex and ChatGPT and can be preloaded with files, packages, skills, and plugins. A self-hosted sandbox means running codex exec-server inside your own environment, where it registers with a restricted key and connects over WebSocket with all traffic outbound. Or you skip the sandbox entirely. Nine partners ship first-class integrations: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel. The harness underneath all of it is the open-source Codex codebase, so developers can read the loop they are paying OpenAI to operate.
Pricing is the most aggressive part of the release. There is no fee for the Agents API itself. Model usage bills at the selected model's normal API rates, with standard charges for OpenAI-provided tools and for sandbox container time on top. A gpt-6-astra session costs what those tokens cost through any other endpoint.
The launch customers supplied the numbers, all of them vendor-reported rather than independently benchmarked. Ciridae said its evaluation score rose from 0.71 to 0.85. “Previously it was pretty cumbersome to observe and orchestrate subagents in our old setup but the new APIs gave us a 4x latency reduction,” said CTO Jack Weissenberger. SafetyKit reported a 60 percent cut in cost per case after migrating a case-review workflow. Hypha, a financial services shop, claimed 86 percent fewer failed agent responses after splitting the harness from the sandbox. Long Lake CTO Rasmus Wissmann framed the division of labor bluntly: “Agents API supplies the harness; the environment, context, and UX stay ours.”
Analysis: the scaffolding layer becomes the product
For two years the agent stack has had a comfortable shape. Model providers sold tokens. A thick middle layer — LangChain and LangGraph, Google's ADK, a hundred orchestration startups — sold the scaffolding that made those tokens behave. That boundary is collapsing from both ends. Anthropic moved first in April with Managed Agents, which bills tokens plus roughly eight cents per session-hour of runtime. Google sells ADK as a code-first framework and then monetizes the runtime through Vertex AI. OpenAI has now gone further by refusing to charge for the layer at all.
Free scaffolding is not charity; it is a commoditization play on someone else's margin. If durable sessions, compaction, and subagent orchestration are table stakes bundled with inference, a framework whose pitch was precisely those features has to find a new one — and the obvious survivors are the ones that stay genuinely model-agnostic or own the tool and auth graph rather than the loop. It also tightens the ratchet on lock-in. The Codex harness is open source and inspectable, but the version OpenAI operates is versioned against OpenAI models and improves with each launch. Porting an agent off it means rebuilding the four things you stopped thinking about.
The constraints are the tell on enterprise readiness. During public beta, Agents API data residency is United States only, and Zero Data Retention is unsupported no matter which sandbox type you pick — including a self-hosted one. That last detail matters more than it looks. Teams that assumed running the sandbox inside their own VPC would satisfy their auditors do not get ZDR anyway, because session state lives on OpenAI's side by design. That is the architectural cost of a managed harness, stated plainly. For a European bank, a hospital network, or anyone under a data-localization mandate, this release is a preview, not a procurement option: the compliance surface is a roadmap item, not a shipped feature.
What to watch next: whether EU residency and ZDR arrive before general availability, how container time actually prices out against Anthropic's per-session-hour meter once real workloads run for days, and whether the nine sandbox partners become a genuine competitive market or a thin veneer over OpenAI's own hosted tier. Watch the middle layer too. The next few weeks of framework release notes will say a great deal about who thinks they still have a business selling the loop.
“Agents API supplies the harness; the environment, context, and UX stay ours.”— Rasmus Wissmann, CTO, Long Lake