OpenAI Agents API: the managed harness, explained
On September 10, 2026, OpenAI moved agent orchestration into its own managed layer. The Agents API puts the harness and infrastructure behind Codex — long-running sessions, sandboxes, tools, subagents — behind a single API call. Here's what it covers, what it deliberately leaves to you, and the four questions that decide whether it's right for production work.
One API call, managed runtime
Before the Agents API, building a custom agent meant assembling the machinery yourself: an agent runtime, context and session management, tools and external data connections, execution environments, and the infrastructure to hold it together. OpenAI already offered pieces — the Responses API for combining models with built-in capabilities like web search, file search, and computer use, and the Agents SDK for defining and orchestrating workflows in your own code.
The Agents API, in public beta since September 10, 2026, goes one level further: the runtime itself becomes a managed service. You specify the task, model, tools, and environment in a single API call, and OpenAI hosts and maintains the underlying harness — the same Codex harness and infrastructure that powers Codex — including context management, tool use, and subagent coordination. Long-running sessions and parallel subagents are part of the managed surface, not something you build.
A harness is the surrounding machinery that lets a model do a job: decide what to try, use a tool, examine the result, keep useful context, and continue. The model supplies capability; the harness determines whether that capability survives contact with your actual files, systems, and edge cases. The Agents API is OpenAI selling you the harness, not just the model.
Where the agent runs is your choice
For the execution environment, you choose between three options: an OpenAI-managed sandbox, your own infrastructure, or a supported third-party sandbox provider — Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel were named at launch. That choice controls the trust boundary: fully managed is simplest, your own VPC keeps data inside your perimeter, and the provider list lets you match sandbox to workload.
File and secret storage plus compute configuration vary by environment choice. This is the first evaluation seam: decide where secrets live before you decide anything else. An agent that needs production credentials is a different security proposition from one working on public data, and the environment choice is where that proposition gets set.
Four questions before production work
The practical question the Agents API raises is how much operational work to hand to a managed runtime versus keep inside your own application and infrastructure. Model capability is the least interesting variable. These four decide it:
1. Permissions
What can the agent touch, and who granted it? A managed runtime that can't express least-privilege tool scopes is a liability no model quality offsets. Map every tool to an explicit grant before the first production run.
2. Observability
Can you see what the agent did, step by step, and replay it? Long-running sessions produce long trails. If the trail isn't inspectable — tool calls, context carried forward, subagent decisions — you can't audit, debug, or defend the outcome.
3. Recovery
What happens when a six-hour session fails at hour five? Managed doesn't mean infallible. Ask how sessions resume, what state survives a failure, and whether retries are explicit or silent.
4. Cost controls
Can you bound spend per task? Long-running agents with tool loops are open-ended by nature. If you can't set a budget the runtime enforces, you're underwriting whatever the agent decides to attempt. Run the cost-per-task math first.
The product is the environment around the model
Step back and the industry direction is unmistakable: the product is increasingly the environment around the model. The engine matters — but so does having a car instead of four wheels and an inspirational keynote. Every major lab shipped dedicated agent infrastructure in September 2026; OpenAI's Agents API is its entry, and DevDay is expected to package this foundation into experiences requiring less assembly: setup, management, distribution, application-building tools.
That convergence is also why the governance layer matters more, not less. A managed harness answers how the agent runs. It doesn't answer whether the agent should have been allowed to act — authorization, tenant isolation, evidence. Those stay your architecture's job. Our managed agents vs. governed missions comparison maps exactly where the managed runtime ends and your governance begins.
Frequently asked questions
What is the OpenAI Agents API?
A managed agent runtime in public beta since September 10, 2026. You specify task, model, tools, and environment in one API call; OpenAI operates the Codex harness underneath — session/context management, tool use, subagent coordination, execution environments.
Where do Agents API agents execute?
Your choice: OpenAI-managed sandbox, your own infrastructure, or supported providers (Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, Vercel). Decide where secrets live before anything else.
Does it replace the Responses API or Agents SDK?
No. Responses API combines models with built-in capabilities; the Agents SDK defines workflows in your code. The Agents API moves the runtime into OpenAI's managed layer — a level above both.
What should I evaluate before production use?
Permissions, observability, recovery, and cost controls — in that order. The operational surface is where managed runtimes win or lose; model capability is table stakes.
From runtime to governance
⚖️ Agents vs missions
Where the managed runtime ends and your governance begins — the evaluation framework.
💰 Cost per task
Bound the spend before you hand a managed runtime an open-ended task.
📰 DevDay 2026
What to expect at the September 29 keynote — confirmed vs. expected.
How this guide is verified
Facts on this page come from OpenAI’s September 10, 2026 Agents API beta launch and the accompanying developer documentation, as independently reported by multiple outlets on launch day (public beta status, the OpenAI-Beta: agents=v1 header, managed sessions/orchestration/context compaction/recovery, your-choice sandbox model, and no separate platform fee beyond model, tool, and container rates). Pricing and access details change; verify against OpenAI’s official docs before building. This page will be refreshed after the September 29, 2026 DevDay keynote with anything OpenAI confirms.