OpenAI Opens the Codex Harness Through the Agents API

OpenAI moved the Agents API into public beta on September 10, 2026, opening the runtime behind Codex and ChatGPT to any developer. The service is managed: it compacts context, coordinates tools, and hands subtasks to subagents. It sits on top of the open-source Codex harness. Neither source reports a platform fee on top of usage.

What you stop building yourself

The agent loop is the part most teams write twice and trust once. MarkTechPost lays the API out as four concepts - Agent, Environment, Session, and the events and items that flow through it - with a session you create, track, resume, and steer while it is still running. That covers the bookkeeping between model output, tool calls, and partial results.

Context and concurrency

As a session approaches its token ceiling, the service compacts context on its own. A tool-search mechanism is meant to cut token spend without invalidating the model cache. Tool calls can be issued programmatically in parallel, and the number of concurrent subagents is configurable; the documented example caps it at three. Supported tool types are MCP servers, custom functions, and built-in tools such as web search.

Where the code actually runs

Execution is not locked to OpenAI. Developers can use the OpenAI-hosted sandbox, self-host through the Codex exec server over WebSocket, run with no sandbox at all, or pick one of nine partner environments: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel. The Decoder singles out Cloudflare, Vercel, and Oracle.

Billing and the two hard limits

MarkTechPost states the billing plainly: no extra fee, with charges for tokens, tools, and container time. The Decoder puts it more narrowly and mentions token usage alone. Two constraints are on the record: data residency is US-only, and Zero Data Retention is not supported. For regulated or EU-bound workloads, that is the line that decides adoption.

Vendor numbers, not benchmarks

The customer figures come from the companies involved, not from independent testing. Ciridae reports an evaluation score moving from 0.71 to 0.85 and a 4x latency reduction; SafetyKit reports 60% lower cost per case; Hypha reports 86% fewer failed responses. Nash.ai says it runs thousands of long-lived agents worldwide. All of it is vendor-supplied.

What neither source pins down

On duration the two reports pull apart: The Decoder describes runs lasting hours, while MarkTechPost frames the service as built for reliable multi-day operation. No hard ceiling appears in either text. Also missing: uptime commitments, an end date for the beta, and any timeline for data residency outside the US.