OpenAI puts its Codex agent harness behind a new Agents API public beta

OpenAI has launched the Agents API in public beta, giving developers access to a managed version of the agent harness and infrastructure used behind Codex. The company says the API lets teams create cloud agents with a single call, then choose where those agents run, including OpenAI hosted sandboxes or supported third party environments.

What Changed

The release turns a large part of agent orchestration into a platform service. OpenAI says the API manages context, tool use, long running sessions and subagent coordination, while developers define the task, model, tools and execution environment. The company is positioning this as a way to move beyond custom prompt chains and bespoke job runners for workflows that need to inspect files, run code, save artifacts and continue across longer spans of work.

OpenAI also introduced hosted sandboxes for developers who want a managed place for agents to work with files, packages, skills and plugins. For teams that cannot use a fully hosted setup, OpenAI lists ecosystem partners including Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel, with options that vary by deployment model, storage, secrets handling, CPU, GPU and memory needs.

Why It Matters

The practical pitch is that agent reliability is becoming an infrastructure problem, not only a model quality problem. OpenAI says the Agents API includes automatic context compaction as sessions approach context limits, tool search to load relevant tools only when needed, and programmatic tool calling for parallel or chained work. Those pieces matter when an agent is expected to operate on a real repository, investigate an incident, or complete a business process that cannot fit inside one short chat exchange.

The announcement includes early customer claims rather than independent benchmarks. SafetyKit says a case review workflow saw a 60 percent reduction in cost per case after moving to the Agents API. Hypha says separating the agent harness from the sandbox reduced failed agent responses by 86 percent. Those numbers are useful signals, but buyers should treat them as workload specific until more public evaluations appear.

What To Watch

OpenAI says the Agents API is available to all developers in public beta, with no separate fee beyond the tokens and tools used. That pricing structure makes experimentation easier, but production use will still require careful limits, observability, sandbox policy and human review for high impact work. CyberOGZ sees the release as a decision point for teams already building agents: keep owning the orchestration layer yourself, or hand more of that responsibility to a model provider and focus engineering effort on tools, data access and approval controls.

Sources

• openai.com

Cover photo by Digital Buggu on Pexels, used under the Pexels License.