OpenAI just gave away the part of your agent you spent six months building
On 10th Sep’26, OpenAI put its Agents API into public beta. It exposes the same harness that runs Codex; the layer that keeps an agent working across long sessions, automatically compacts earlier context as the window fills, loads tool definitions only when they're needed, runs subagents in parallel, and recovers when a step fails. You supply the tools and choose where the code runs: an OpenAI-hosted sandbox, your own infrastructure, or one of nine partner environments (Cloudflare, Vercel, Modal, E2B, Daytona, DigitalOcean, Oracle, Blaxel, Runloop). The harness is the open-source Codex codebase, so you can read what it does. OpenAI's launch page carries customer figures: SafetyKit reports a 60% reduction in cost per case after migrating its case-review workflow; Hypha reports 86% fewer failed agent responses from separating the harness from the sandbox. Those are customers' numbers, published by the vendor.
What it actually means
For two years the real work of building with AI hasn't been the model, but the scaffolding round it. Context that runs out halfway through a long job. Keeping a session alive for hours. Stopping two subagents trampling each other's work. Most teams built that themselves or bolted on a framework, and nobody ever thought it was the interesting part. OpenAI has now made that layer free, maintained, and versioned to each model release.
Free isn't free, it's coupled: OpenAI says the harness improves alongside its models, which is another way of saying it's tuned to them, and the layer you'd be handing over is precisely the one that used to make you portable. Keep the sandbox somewhere you control, keep your tools defined in MCP, which isn't theirs, and the harness is a component you slotted in. Skip that and it's the floor you're standing on.
The orchestration layer was never your product.
60%cut in cost per case one customer reports after moving to the managed harness
Source: OpenAI · Friday 11 September 2026