Running agents for days? OpenAI now rents the harness. What we'd test first

5 min read

Read this if you: the builder deciding whether to run long, unattended agent jobs on their own harness or rent someone else's, who wants to know what OpenAI's Agents API actually offers as of this week and what a reviewer here would put it through before trusting it with a dependency.

On 10 September 2026 OpenAI's developer changelog recorded a one-line release with a long shadow: "Released the Agents API in public beta. Build agents with a managed Codex harness while OpenAI handles session orchestration, context compaction, and recovery." A developer specifies a task, a model, tools and a compute environment; the work runs in OpenAI-hosted sandboxes, in the developer's own infrastructure, or in partner environments from Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel. OpenAI says there are no additional fees for the API itself; you pay standard rates for the tokens and tools consumed.

What it changes for the reader of this site: the harness that runs an agent unattended for days, the thing every team has been half-building, is now a rented commodity. That moves the review question. It's no longer "can I keep a session alive"; it's "what does the rented harness do to the services my agent calls, and what does it do when they fail".

What shipped, as documented

The features listed are durable sessions with progress streaming, automatic context compaction, tool search to cut token use, parallel tool calling, subagent delegation, and integration of custom tools or MCP servers, the standard way an agent is handed a tool. One early user, Jack Weissenberger, chief technology officer at Ciridae, said as quoted by Investing.com that an "evaluation score increased to 0.85 from 0.71 and produced a 4x latency reduction for subagent workflows." The New Stack reports that OpenAI's own research organisation logged 3.1 agent-workdays per human workday by mid-August, which is the scale the product is built for.

What the documentation we read does not state: a service-level agreement, rate limits, or caps on session duration. For a harness whose whole pitch is running for days, those three are the terms that matter, and a reviewer would want them in writing before the first paid run.

What a reviewer here would test first

No verdict without a run; that's the rule. But the run we'd design is clear enough to describe.

If you're deciding now, before a review exists: rent the harness for jobs whose failures are cheap and visible, keep your own harness for jobs that spend money against third parties, and write the retry and idempotency rules into your tools rather than trusting the runtime to know a paid call from a free one.

The release in one table

ItemDetail
Date10 September 2026, public beta (OpenAI developer changelog)
WhatA managed Codex harness: session orchestration, context compaction, recovery handled by OpenAI
Where it runsOpenAI-hosted sandboxes, your own infrastructure, or Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, Vercel
Features listedDurable sessions with progress streaming, automatic compaction, tool search, parallel tool calls, subagent delegation, custom tools and MCP servers
PriceNo additional fee for the API; standard token and tool rates apply
Not documented in what we readSLA, rate limits, session duration caps

What we haven't run

We haven't run a job on it; this is a dispatch, not a review, and it carries no verdict. OpenAI's announcement post was not reachable from our tools, so the feature list is from the changelog and the reporting linked below, and the Ciridae figures are as Investing.com quoted them. A widely repeated "$7,000 a day" cost figure appears in a headline we could not verify against body text, so it isn't here.

Questions builders ask

Is the Agents API free?

The API carries no extra fee, per OpenAI. The tokens and tools your agent consumes are billed at standard rates, and a days-long session consumes a lot of both.

Can I run it on my own servers?

Yes, per the documentation: your own infrastructure is one of the three compute options, alongside OpenAI's sandboxes and the named partner environments.

Would Agent Reviews recommend it?

Not yet. A recommendation here follows a run. The tests above are the run we'd do.

Sources