Skip to content
NewsProduct launch

OpenAI opens Agents API in public beta, making the Codex harness available to all developers

· by Pondero Newsdesk

The short version

OpenAI made its Agents API available in public beta on September 10, 2026, giving any developer access to the managed orchestration layer behind Codex with no additional API fee beyond model tokens and paid tools.

OpenAI opens Agents API in public beta, making the Codex harness available to all developers

Codex runs on a managed harness. As of September 10, 2026, that harness is a public API. OpenAI opened its Agents API in public beta, giving any developer with an API key access to the same session management, context compaction, and multi-agent orchestration that powers its own coding assistant. No additional API fee applies beyond model tokens, tool calls, and container compute time.

What

The Agents API wraps the Codex harness into a general-purpose endpoint. Developers supply tools and choose an execution environment; OpenAI's infrastructure handles sessions, context window compaction, and recovery when limits approach.

Capabilities include sandbox code execution, file editing, artifact generation, MCP server connections, and delegation to subagents with independent contexts, per MarkTechPost's coverage of the launch. A built-in tool search mechanism reduces token overhead by fetching only the relevant tools during planning rather than passing the full list each call.

Three execution modes are available. First: an OpenAI-managed sandbox, configurable with files, packages, and skills. Second: a self-hosted option using codex exec-server over a WebSocket connection. Third: one of nine partner sandboxes at launch (Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel).

Why it matters

For developers building agentic workloads, the Agents API is the practical successor to the Assistants API. Sessions, context limits, and state management were previously manual; the new API absorbs them at the platform level.

OpenAI disclosed early adopter metrics at launch. Read these as selected customer results, not typical outcomes: Ciridae reported evaluation scores rising from 0.71 to 0.85 with 4x lower latency; SafetyKit reported 60% lower cost per case; Hypha saw 86% fewer failed responses after migrating, per MarkTechPost.

Two current limitations matter for enterprise buyers. Data residency is US-only, and Zero Data Retention is not supported, which blocks regulated-data use cases outright until OpenAI addresses those gaps.

Competitively, this puts OpenAI's managed harness directly against Anthropic's Agent SDK and Google's ADK. The key tradeoff is portability: OpenAI's approach reduces wiring by managing state and context server-side; Anthropic and Google's SDKs run inside your own stack with more control and more integration work required.

What to watch next

Uptake among developers currently on the Responses API is the clearest early signal of whether the managed harness is worth the platform dependency. Nine partner sandboxes at launch suggests OpenAI expects workloads to spread across cloud providers rather than consolidate in a single environment. Progress on ZDR support and non-US residency will determine how quickly regulated enterprise buyers can seriously evaluate this option.

Sources