OpenAI Launches Agents API, Opening Codex’s Engine to Developers

OpenAI opened its Agents API in public beta on September 10, 2026, giving any developer direct access to the managed “Codex harness” — the same orchestration and infrastructure layer that powers Codex and ChatGPT Work — behind a single API call. The release, announced the same morning as the company’s GPT-Live-1 voice model going generally available in the API, marks a shift from OpenAI selling chat completions to OpenAI selling the machinery needed to run autonomous, long-running AI agents.

OpenAI CEO Sam Altman. Photo Credit: Steve Jurvetson, via Wikimedia Commons, licensed under CC BY 2.0

For developers who have spent the past two years hand-building agent orchestration, memory management, and sandboxing from scratch, the announcement effectively productizes the hardest, least visible part of building a reliable AI agent.

What Shipped on September 10

According to OpenAI’s official announcement, “Introducing the Agents API,” the new managed service lets developers create a production-ready agent with a single API call by specifying the task, model, tools, and execution environment. OpenAI runs the agent loop — coordinating model calls, tool use, and context — on its own infrastructure, while developers choose where the agent’s actual computation happens.

Under the hood, developers work with four building blocks. An Agent defines what the model has access to — its instructions, tools, and any connected MCP servers. An Environment, which is optional, is where the agent’s compute actually runs. A Session keeps a given piece of work alive across many back-and-forth turns instead of resetting each time. And a running session continuously reports back what it’s doing through a stream of events the calling application can follow in real time. Within any given session, an agent can execute code, edit files, pull in web search results, invoke pre-built skills, generate finished output, and split a task across multiple subagents working within a capped concurrency limit.

Sandboxes: Bring Your Own, or Let OpenAI Manage It

Teams that want the fastest path to a working agent can lean on a sandbox OpenAI hosts itself — built on the identical sandboxing setup already running Codex and ChatGPT in production, with OpenAI handling provisioning and management while developers simply supply the files, packages, skills, and plugins their agent needs. Developers with stricter requirements can instead run agents on their own infrastructure or route them through one of nine named partner sandbox providers — Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel.

That partner list is notable in its own right: it puts OpenAI in the position of routing developer workloads to infrastructure competitors, including Oracle, which is simultaneously one of OpenAI’s own largest compute suppliers.

Two limitations are worth flagging for teams evaluating the beta: data residency is currently US-only, and Zero Data Retention is not yet supported — both of which will matter for regulated industries considering production use.

Context Management and Cost

To support agents that need to keep working for hours or days, OpenAI has built automatic context compaction into the service: as a session approaches its context limit, the API compacts earlier context to preserve the information an agent needs to keep going, without developers having to implement that logic themselves. A related feature, tool search, loads relevant tool definitions only as needed, which OpenAI says reduces token usage and cost while preserving model caching. Once tools are loaded, programmatic tool calling lets agents run calls in parallel and chain or filter results in code.

There is no additional platform fee for using the Agents API itself — developers pay only for the tokens, tools, and container time their agents consume, according to OpenAI’s pricing page.

Early Adopters and Reported Results

OpenAI’s rollout post lists several companies already running live workloads on the new API, among them Ciridae, Long Lake, WithCoverage, SafetyKit, Dwelly, Hypha, deepsense.ai, and Nash.ai. Two of those results stand out in OpenAI’s telling: Ciridae says its latency dropped by roughly four times once it started using the built-in subagent coordination, and Hypha says its failure rate on responses fell by 86% after switching to the managed harness. Both figures come directly from OpenAI’s own announcement and have not been checked by an outside party.

Why This Matters

For founders and independent developers, the practical effect is a shorter path from prototype to production: instead of stitching together a custom orchestration loop, a third-party sandbox provider, and hand-rolled context management, teams can now call a single managed endpoint that runs the same harness OpenAI uses internally for Codex. That doesn’t eliminate the engineering work of building a good agent, but it removes a substantial amount of undifferentiated infrastructure work that most teams were previously duplicating.

For enterprises already using ChatGPT Work, the release tightens the relationship between OpenAI’s consumer-facing agent products and the raw infrastructure behind them — capabilities that were previously locked inside Codex are now versioned and exposed through a public API, meaning updates to the underlying harness should reach developers on a similar cadence to OpenAI’s own products.

The launch also lands against a backdrop of heightened scrutiny of autonomous AI agents industry-wide. Concerns about agent swarms acting outside their intended scope have featured prominently in recent industry debate over how fast AI capabilities should advance, and any infrastructure that makes it easier to run large numbers of coordinating agents in production will likely draw continued attention on the safety and governance side, even as it lowers the technical barrier for legitimate use cases.

What Could Happen Next

• Broader data residency and Zero Data Retention support. Both are common requirements for enterprise and regulated customers, and their absence in the beta will likely limit adoption until OpenAI addresses them.
• Competitive responses from Anthropic and Google. Both companies offer their own agent-building tools; whether either moves to expose an equivalent managed harness is worth watching.
• Independent verification of the reported customer results. The latency and reliability figures cited in OpenAI’s announcement come from the company’s own customers and have not been independently benchmarked.
• How pricing evolves out of beta. “No additional fee beyond tokens and tools” is the current beta pricing model; whether that structure holds at general availability is unconfirmed.

Conclusion

The Agents API is less a new product than an admission of what building reliable AI agents actually requires: not just a capable model, but a harness that manages context, coordinates tools, and keeps sessions alive for hours or days. By exposing the same infrastructure that runs Codex and ChatGPT Work, OpenAI is betting that most developers would rather rent that machinery than rebuild it — a bet that, if it pays off, could make the Agents API as consequential to OpenAI’s developer business as the original ChatGPT API was to its consumer business.

FAQ

What is the OpenAI Agents API?
It’s a managed API, released in public beta on September 10, 2026, that gives developers direct access to the same “Codex harness” infrastructure OpenAI uses internally to run Codex and ChatGPT Work — handling session orchestration, context management, tool use, and subagent coordination.

How much does it cost?
There is no additional platform fee for the API itself in the current beta; developers pay for the tokens, tools, and container/compute time their agents actually use, per OpenAI’s pricing page.

Can I use my own infrastructure instead of OpenAI’s sandboxes?
Yes. Developers can run agents on OpenAI-hosted sandboxes, their own infrastructure, or through nine named partner sandbox providers, including Cloudflare, DigitalOcean, and Vercel.

Is the Agents API ready for regulated industries?
Not fully yet. The current beta is US-only for data residency and does not support Zero Data Retention, both of which are typically required for healthcare, finance, and government workloads.

Sources

• OpenAI, “Introducing the Agents API” — openai.com
• OpenAI Developer Community, “Introducing the Agents API and hosted sandboxes” — community.openai.com
• MarkTechPost, “OpenAI Launches the Agents API in Public Beta, Putting the Codex Harness Behind One API Call” — marktechpost.com
• OpenAI Release Notes — openai.com