OpenAI Agents API Public Beta: The Codex Harness Is Now a Product

OpenAI spent the first week of September shipping one thing after another — GPT-6 Astra on the 3rd, two Codex CLI updates, then on September 10: the Agents API in public beta, GPT-Live-1 in the API, and ChatGPT for Financial Services. Same morning, eight design partners quoted on the launch post.

The through-line is pretty clear if you read the posts back-to-back. OpenAI is moving the Codex harness — the thing that actually runs the agent, manages context, calls tools, coordinates subagents, and keeps it from falling over after six hours — from something developers run themselves to something they call with an API.

"Build and run cloud agents with the Codex harness, fully managed by OpenAI." That's the subtitle. The rest is detail.

What the Agents API Actually Is

A single API call creates a production-ready agent session. You specify the model (the launch samples use gpt-6-astra), the task, the tools, and the environment. OpenAI runs the harness. You pick where the agent computes.

The harness does the stuff every agent builder has been writing by hand: session management, orchestration, context compaction when a session approaches its limit, tool search that loads relevant tool definitions only when needed, programmatic tool calling so agents can run calls in parallel and filter results in code, subagent delegation where the main agent breaks a task into pieces and assigns them to subagents that each keep their own context, and recovery so a session can resume where it left off.

OpenAI's own words from the launch post, on why this exists: "Useful agents need a powerful harness that manages context, uses tools efficiently, and coordinates subagents." They've also said they learned this scaling Codex and ChatGPT for Work to millions of people — the harness is the thing they had to get right to make long-running agents actually work, and now they're selling it.

There's no separate fee for the API. You pay for model tokens at the model's rates, OpenAI tools at standard rates, and OpenAI-hosted sandboxes at container rates. Self-hosted or partner environments bill through their own providers.

The Environment Choice: OpenAI Sandbox, Your Infrastructure, or Nine Partners

This is the piece that matters for deployment. Three options:

OpenAI-hosted sandbox — same infrastructure behind Codex and ChatGPT. Configured with your files, packages, skills, and plugins. Get started fast, scale without managing it.

Your own infrastructure — self-hosted. You choose the execution environment and provide the tools. Full control, full responsibility.

Partner sandboxes — nine named partners: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel. Managed environments or deployments inside your VPC, with specific file and secret storage, and different CPU, GPU, and memory configurations.

The partner list is notable. Cloudflare and Vercel on the edge. Modal and DigitalOcean for compute density. Oracle for enterprise. E2B and Daytona for developer environments. This isn't a random partner dump — it covers the real deployment topologies teams actually use.

What the Design Partners Reported

OpenAI quoted eight design partners on the launch post. A few specifics worth noting:

Ciridae (Jack Weissenberger, CTO): evaluation score went from 0.71 to 0.85, subagent flows gave a 4x latency reduction, and the orchestration used to be "pretty cumbersome" to observe and manage in their old setup.

SafetyKit (Bhavyansh Sabharwal, MTS): 60% reduction in cost per case after migrating their case review workflow, with lower latency and improved token efficiency.

Hypha (Serhii Shchoholiev, Lead Engineer): in financial services, separating the agent harness from the sandbox reduced failed agent responses by 86%. "Earning customers' trust is critical in financial services," he wrote.

Dwelly (Dmitry Khanukov, Co-founder & CTO): could fan out work across hundreds of agents, run them asynchronously, and collect results later without keeping infrastructure idle between peaks — "bursty workloads" handled naturally.

Nash.ai (Aziz Alghunaim, Co-founder & CTO): deploys thousands of long-running agents managing hundreds of millions of deliveries across global logistics. The API gives them "the durable session and orchestration layer we need for agents operating continuously in production."

These are vendor-chosen quotes, not independent benchmarks. Treat them as directional. But the consistency is interesting — multiple teams reporting that the harness layer was the thing they were spending engineering time on, and the API removes it.

The Open-Source Angle

The Agents API is powered by the open-source Codex harness. OpenAI operates and maintains it; developers can inspect and learn from the public codebase on GitHub. The API provides versioned access to harness capabilities with each model launch, so a new model arrives with the harness already tuned for it.

That's the pitch: you no longer chase the harness. OpenAI maintains it alongside the models, and each launch ships with harness improvements already in place.

Recent harness improvements cited in the launch post include context compaction that preserves information the agent needs as a session grows, tool search that reduces token usage by loading tool definitions only when needed, programmatic tool calling for parallel and chained operations, and multi-agent support with subagents that maintain their own context.

What's Not There Yet

A few limits worth knowing:

• Data residency is currently US-only.
• Zero Data Retention is not supported, even when execution is self-hosted.
• The launch samples name gpt-6-astra; the supported model list beyond that is whatever the docs say on the day you build.
• It's public beta. The post says OpenAI will "iterate quickly based on your feedback as we work toward general availability" — GA, rate limits, session-length limits, and any container-rate table specific to Agents API sandboxes are all still to come.

The Week the Harness Moved

Context helps here. September 3: GPT-6 Astra and Codex CLI 0.153.0. September 9: Codex CLI 0.154.0. September 10: Agents API public beta, GPT-Live-1 in the API at $0.05/minute with 0.8-second turn-taking latency, and ChatGPT for Financial Services with Morgan Stanley and Evercore as design partners.

Three Codex-related releases in eight days, then the harness itself becomes a product. If you've been building agent infrastructure, the question this week is whether you're still writing harness code or whether you should be calling an API instead.

Why This Matters Beyond "Another API"

The agent space has a real split right now. On one side: model capabilities — GPT-6 Astra, Claude Fable 5.1, the benchmarks, the scores. On the other side: the stuff that actually makes an agent work in production — session management, context handling, tool orchestration, recovery when it falls over, cost control across long runs.

Most of the public conversation is on the model side. The Agents API is OpenAI's bet that the harness side is equally important and that they've already solved it well enough to sell. The design-partner quotes line up with that — multiple teams saying the harness was the engineering problem, not the model.

For developers, the practical question is straightforward: if you're running Codex or building your own agent harness, does the managed version give you enough to switch? The answer depends on your deployment constraints (data residency, ZDR, which sandbox you need) and on whether you trust OpenAI to maintain the harness versioned to each model launch.

The API doesn't cost extra. The cost is the deployment constraints and the control you give up. That's the tradeoff, and it's a real one.

Published Tuesday, September 15, 2026. Sources: OpenAI's "Introducing the Agents API" launch post (September 10, 2026); OpenAI Agents API overview and pricing pages in the developer docs; CellCog's launch-day analysis; AI Weekly and MarkTechPost reporting. Follow AIPress for ongoing coverage of AGI, ChatGPT, Claude, OpenAI, Anthropic, and the AI industry.