OpenAI Agents API: The Agent Loop Is Now a Free Hosted Product

OpenAI's Agents API is a managed service for building and running cloud agents, and it went into public beta on 10 September 2026. The OpenAI Agents API is, in effect, the agent loop as a hosted product. In plain terms, you write the tools and the instructions, and OpenAI runs the loop: the session, the context, the recovery when something breaks. The orchestration you used to build yourself is now someone else's product.

That sounds like a developer story. It is a marketing story, because the thing being industrialised here is software that visits websites, reads them and acts on them without a person watching.

What OpenAI actually shipped on 10 September 2026

The Agents API is powered by what OpenAI calls the Codex harness, the same scaffolding behind its coding agent. It gives developers durable sessions that carry work across many turns, streaming progress so you can watch a job run, and connections to custom tools and MCP servers. MCP, the Model Context Protocol, is the open standard for plugging a model into an external tool or data source.

Four capabilities matter more than the rest:

• Long-running sessions. An agent can keep working across a task that takes far longer than a single request.
• Context compaction. The service summarises earlier work so a long job does not run out of room. I have written before about how the context window grew 488 times while the search box did not, and compaction is what happens when even a large window is not enough.
• Subagent coordination. One agent can hand work to others.
• Sandboxed execution. Code runs either in an OpenAI-hosted sandbox or in a compute environment the developer nominates.

The API changelog carries the technical detail. The pricing is the part I would put in front of a CFO. OpenAI charges no separate fee for the Agents API. You pay for the models and tools each session consumes. There is no orchestration line item.

Why a zero-fee orchestration layer changes the volume of agents

Removing a fee is not a discount. It is a decision about where the market forms.

Until now, building a reliable agent meant building the boring parts yourself: retry logic, state that survives a crash, context that does not overflow, a safe place to run code. That work was the barrier to entry, and the barrier is what kept the number of production agents small.

Price that work at zero and hand it to anyone with an API key, and you should expect the population of agents pointed at the open web to grow quickly. Not because agents suddenly became more useful, but because they became much cheaper to attempt.

This is the same mechanism I described in what 261 model releases show about shelf life: when the price of adequate collapses, the thing gets deployed everywhere the marginal cost rounds to nothing.

What this does to your traffic, your analytics and your funnel

Here is the connection to search that most coverage of the OpenAI Agents API will skip.

An agent that browses is a visitor your analytics was never designed for. It does not scroll. It does not dwell. It does not hover, hesitate or come back tomorrow. It arrives, extracts, and leaves. Every engagement signal you use as a proxy for quality goes flat, and it goes flat on traffic that may be doing real commercial work on a buyer's behalf.

I set out the mechanics of that in how AI browsers rewrite customer acquisition. The Agents API does not introduce the problem. It industrialises it, by making the agent the default unit of software rather than an experiment.

Three consequences worth planning for now:

Your bounce rate becomes partly fictional. If a rising share of sessions are agents completing a task in one visit, average session duration falls while the business outcome improves. Anyone reading those two numbers together without splitting agent traffic out will reach the wrong conclusion and, worse, act on it.

Your content gets read by something that cannot be persuaded. An agent does not respond to a hero image, a testimonial carousel or social proof. It responds to facts it can extract and check. That is the same format instruction the citation data has been giving for a year, arriving now through a second door.

Your forms become the choke point. An agent acting for a buyer can read your pricing page. It cannot ring your sales team. Anything you gate is invisible to the layer that increasingly does the shortlisting.

The single-provider risk nobody prices

There is a governance point here that deserves more than a footnote.

A managed harness means OpenAI owns your agent's control loop. When the orchestration, the model, the sandbox and the tool protocol all come from one vendor, the swap cost is no longer a model name in a config file. It is your entire runtime.

That is worth holding next to what the release data says about durability. Across a workbook of 261 significant model releases from 15 labs, 75.5% are already superseded or retired, and the median model on sale is around 60 days old. Building on a managed layer from a single provider, in a market that replaces its parts every couple of months, is a decision to accept a particular kind of risk rather than an absence of one.

Ask the question you would ask of any vendor: what happens to this system if the provider deprecates the harness, changes the pricing basis, or restricts your use case? If nobody can answer, that is the answer.

What to do about the OpenAI Agents API this quarter

Four things, none of which need new software.

Split agent traffic out of your analytics now. Not later, when the volume is obvious. You cannot prove a change you never baselined, and the baseline is only available before the shift, not after.

Publish the facts an agent needs. Prices, specifications, availability, limits, with dates on them. A page that says "contact us" is a page an agent cannot use and will route around.

Check your crawler policy against your commercial policy. Many sites block automated access by reflex. If agents are increasingly acting for buyers, a blanket block is a decision to be absent from the shortlist, and it is usually made by someone who was never told that was the trade.

Write the swap clause before you build. Abstract the provider behind your own interface on day one. The cost of doing that at the start is a week. The cost of doing it after a deprecation notice is a quarter.

The OpenAI Agents API is not interesting because agents are new. It is interesting because the hard part of running them just became free, and the web is about to be read by a great deal more software that does not behave like a person. Measure that before it arrives, not after.