Agent Horizon

OpenAI has launched the Agents API in public beta, making the same Codex harness that powers its coding agent available to any developer through a single managed API. The service handles orchestration, long-running sessions, context compaction, and recovery, while developers supply their own tools and choose where the agent’s code actually executes: an OpenAI-hosted sandbox, their own infrastructure, or one of several partner sandboxes including Cloudflare, DigitalOcean, and Oracle.

This is a plumbing announcement rather than a flashy new model, but plumbing is often where the real bottleneck sits. Building a reliable agent that can run for hours or days, recover from crashes, manage a shrinking context window, and coordinate subagents is genuinely hard engineering work that most teams end up rebuilding from scratch. By packaging that infrastructure as a managed service, OpenAI is betting that developers would rather use a known system than write their own orchestration layer from zero.

It’s also a competitive move worth noting in context: Anthropic has its own Claude Agent SDK, and Google has its Agent Development Kit, so this puts OpenAI’s agent infrastructure on more equal footing as a standalone product rather than something bundled only inside ChatGPT or Codex itself. Early adopters cited in OpenAI’s announcement report improvements in agent reliability and latency after switching, though those figures come from individual customers and shouldn’t be read as universal benchmarks.

The practical caveat for teams considering this: it’s a beta, pricing follows standard token and sandbox-provider rates rather than a flat fee, and features like Zero Data Retention aren’t yet supported, so anyone with strict compliance requirements should read the fine print before committing production workloads to it.