# OpenAI Opens the Codex Harness as a Managed Agents API

The public beta charges no platform fee beyond tokens and tools, and lets developers run agents in OpenAI sandboxes or on Cloudflare, DigitalOcean and Oracle infrastructure.

- Published: 2026-09-11T06:22:22.078Z
- Canonical: https://polylog.news/ai/2026-09-11/openai-opens-the-codex-harness-as-a-managed-agents-api
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [OpenAI News](https://openai.com/index/introducing-the-agents-api), [OpenAI Developer Docs](https://developers.openai.com/api/docs/guides/agents-api/overview), [MarkTechPost](https://www.marktechpost.com/2026/09/10/openai-launches-the-agents-api-in-public-beta-putting-the-codex-harness-behind-one-api-call/)

OpenAI has turned the runtime behind Codex into a public application programming interface (API). The [Agents API](https://openai.com/index/introducing-the-agents-api), in public beta since September 10, exposes the Codex harness as a managed service. OpenAI runs session state, orchestration, context compaction and recovery, while the calling application supplies the tools and chooses where the code actually executes.

The division of labor is the point. Most teams building long-running agents have written the same supporting code themselves: retry logic, context trimming as a session outgrows the context window, and checkpointing so a crashed run does not have to restart from zero. OpenAI's [developer documentation](https://developers.openai.com/api/docs/guides/agents-api/overview) moves that layer onto its servers. Execution environments can be OpenAI-managed sandboxes or third-party infrastructure, with launch integrations from Cloudflare, DigitalOcean and Oracle.

Pricing is the aggressive part. OpenAI charges no separate platform fee for the Agents API and bills only for tokens and tools consumed, undercutting every orchestration vendor that charges for this supporting code. The company cites early users reporting four times lower latency, 60 percent lower cost per task and 86 percent fewer failed responses, figures [MarkTechPost repeats](https://www.marktechpost.com/2026/09/10/openai-launches-the-agents-api-in-public-beta-putting-the-codex-harness-behind-one-api-call/) from the announcement. Those numbers come from OpenAI, describe unnamed workloads measured against unnamed baselines, and have not been independently reproduced. They should be treated as marketing claims until someone publishes an independent comparison.

The strategic significance does not depend on those numbers. Whoever controls the agent loop controls the session state, the tool registry and the context policy, and that is a far more durable position than simply serving completions, because customers who build on that session state face high costs to switch providers.

## What this means

Framework vendors that sell orchestration, session persistence and context management now compete with a free layer bundled into the model provider's API, and their remaining defensible ground is model-neutrality for buyers who refuse single-vendor lock-in. Cloud partners named at launch gain a distribution channel for sandboxed compute. The cost to developers is a harder migration later, because session state and compaction policy live on OpenAI's side rather than in portable application code.

## What to watch

- Whether Anthropic and Google respond with comparable managed agent runtimes, which would confirm that the orchestration layer is being absorbed into the model API across the industry.
- Whether independent benchmarks reproduce anything close to the claimed 60 percent cost reduction per task, since vendor-reported efficiency gains often shrink under neutral testing.
- How agent frameworks that sell portability across model providers position themselves in the next quarter, which will show whether multi-vendor support is still worth paying for.