Do you still need to host your own agent harness in 2026?

What actually changed on September 10, 2026?

OpenAI took the harness that runs Codex and sold it as a service. The Agents API is in public beta for every developer, and one call gets you long-running sessions, orchestration, tool use, subagents and automatic context compaction as the session nears its limit. OpenAI's documentation draws the comparison itself: "A hand-built loop makes your code responsible for session state, context compaction, crash recovery and subagent coordination." Those four things are where an in-house harness spends its maintenance budget.

Pricing is the part that moves a decision. There is no additional fee for the Agents API itself. You pay model tokens at the selected model's normal API rate, OpenAI's built-in tools at their standard rates, and container time on whichever sandbox runs the work. The announcement page blocks automated fetches, so every figure on this page is cross-checked against OpenAI's developer documentation and the announcement thread on its own developer forum, both dated September 10, 2026.

Sandboxes are what make this a genuine build-versus-buy question instead of a vendor pitch. Compute runs in an OpenAI-managed sandbox, in one you host yourself, or on one of nine partners with first-class integrations: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel. Adopting the harness does not require handing over the execution environment.

Nine days ago, maybe worth building argued that building a general-purpose agent harness was not worth it, resting the case on one open-source harness with 245,800 stars. OpenAI has now made the same argument from the commercial side, with a product instead of a repo.

What does the managed harness cost, and where does the bill land?

The harness is free. The container is not, and that is where teams get surprised.

Anthropic's comparable product gives you the shape of the number. Claude Managed Agents went to public beta on April 8, 2026, bundling agent loop, tool execution, sandbox container and state persistence behind REST endpoints, billed at standard token rates plus $0.08 per session-hour of active runtime. OpenAI has not put a flat session-hour price on its own sandbox, it charges standard container rates instead, but the line item is the same one. An agent sitting idle inside an open session is still billing.

That warning came from inside the announcement thread, not from a critic. VeitB, posting in OpenAI's developer forum on September 10, 2026, wrote: "PSA: If you are new to using hosted sandboxes, please be careful when spinning up new containers and calculate the costs first. I learned this the hard way!" Worth reading next to our numbers on what it costs to run coding agents full time, where task length and concurrency already turned out to be the levers that move a bill, not model choice.

Money is also flowing into that container layer, which tells you it is not a rounding error. Daytona, one of the nine named partners, raised a $24M Series A in February 2026 led by FirstMark Capital, building fast-starting sandboxes for agents to execute code in. The container layer has its own vendors, its own pricing pages and its own venture funding, so budget it as a real line rather than an afterthought.

Does self-hosting the sandbox solve your compliance problem?

No, and this is the sharpest thing in the documentation. OpenAI states that "The Agents API currently supports data residency only in the United States and does not support Zero Data Retention (ZDR)," and then closes the obvious escape hatch in the next breath: "Choosing a self-hosted sandbox does not make the Agents API ZDR-eligible."

The second sentence is the one that catches people. Self-hosting the sandbox moves where the agent's code executes. It does not move where the session lives. Orchestration, compaction, recovery and the retained conversation state all sit with OpenAI regardless of whose machine ran the tool call. If you hold a ZDR contract or an EU residency requirement, the self-host option does not rescue you during the beta, and no amount of network configuration on your side changes that.

This is the condition that decides the most cases, and it is not a philosophical objection to managed services. It is a clause somebody in your company already signed. If you are in healthcare, finance, or anything EU-facing under a data processing agreement, you keep your own harness and the rest of this page is academic for you.

Is programmatic tool calling worth switching for on its own?

For data-heavy agents, close to it. This is the one feature in the launch that changes your bill rather than your convenience, and it is the strongest technical reason to look at the managed harness even if you were happy with your own.

Rather than every tool result landing in the context window, the model writes JavaScript that coordinates your tools, and that program runs in a fresh isolated V8 runtime. It can call tools in parallel, loop, branch, hold intermediate results in the runtime, and filter or combine them in code, so only the relevant slice ever comes back into context. The runtime is deliberately bare: no Node.js, no package installation, no direct network access, no general-purpose filesystem, no subprocess execution, no console, no persistence. Your tools remain the only path to the outside world, so the security boundary is the one you already reason about with ordinary function calling.

A tool call that returns 10,000 rows and gets filtered to twelve inside the runtime is billed as twelve rows, once, instead of 10,000 rows on the way in and again on every turn that carries the history forward. Pair that with tool search, which OpenAI says "loads tool definitions only when needed," and the token floor for a wide toolset drops before you have touched a model or a prompt. That is a direct cost lever, and rebuilding it yourself means writing and securing your own sandboxed code interpreter.

OpenAI's launch cites three customers: Ciridae moving an evaluation score from 0.71 to 0.85 with a 4x latency reduction, SafetyKit reporting 60% lower cost per case after migrating, and Hypha reporting 86% fewer failed agent responses. Those are vendor-supplied and self-selected, so treat them as the direction of travel and not as your forecast. The V8 runtime mechanism is verifiable in the docs. The customer numbers are not.

When do you still run your own harness?

Six conditions, and any one of them is enough on its own.

• Zero Data Retention or non-US residency. Settled above. The self-hosted sandbox does not buy either during the public beta.
• Hard egress control. If your security model requires that you can name every destination an agent's traffic reaches and hold the logs, a managed orchestration layer you cannot inspect is a gap you will have to explain to an auditor.
• Custom sandbox requirements. Specialized hardware, a bespoke base image, GPUs attached to the agent's own runtime, or a filesystem layout the API's environment model does not express. Self-hosting covers a lot of this, but check your exact case against the environment schema before you assume.
• A model this harness will not run. The Agents API runs OpenAI's models. If your production loop depends on an open-weights model you host yourself, or on mixing vendors inside a single run, the harness does not fit the workload.
• Sustained high volume. Container time is metered. At low volume that is noise. At sustained volume with long-lived sessions it becomes a real line, and your own infrastructure amortizes in a way rented containers never will.
• Single-vendor dependency on the core loop. If the agent loop is the product, putting it inside one vendor's service is a different bet than building your workflow on one AI vendor, and a heavier one.

Notice what is missing. "We want control" is not on the list, because control over the sandbox is now available without writing the harness, and control over the session data is not available either way.

How heavy is the lock-in, actually?

Heavier than model lock-in, and this is the caveat that belongs on every recommendation to adopt the managed harness.

Changing models is a string and a rerun of your evals. Changing harnesses is a rewrite, because the harness is where your orchestration logic lives. Once OpenAI owns your session state, your compaction policy, your crash recovery and your subagent coordination, those things are no longer in your repository. They are configuration against someone else's service, and the migration path back is rebuilding what you deleted.

Both products are also in public beta. OpenAI's since September 10, 2026, Anthropic's since April 8, 2026. Beta means the surface can change under you, and two vendors selling the same shape does not make them interchangeable: the endpoints, the environment model and the subagent configuration differ, much as agent skills turned out to port only partially between Claude and Codex.

The mitigation is boring and it works. Keep your tool definitions, your eval set and your prompts in your own repository, structured so the harness is the thing you call rather than the thing you live inside. Then the managed harness is an implementation detail you can swap in a sprint instead of a quarter.

What test should you run before migrating?

Pick one workload you already have eval scores for. Not a demo, not a greenfield service, something with a baseline you trust.

Run it on the Agents API with a self-hosted sandbox, so you are testing the harness rather than testing the harness and a new execution environment at the same time. Then compare three numbers against your current loop: pass rate on your existing eval set, total cost including container time, and p95 latency. If the managed harness does not win on your evals, the thing you bought was somebody else's maintenance schedule, and you already have one of those.

If it does win, migrate the workloads that clear all six conditions above and leave the rest on your own harness. Running regulated work in-house and everything else on the managed API is a reasonable end state, and it survives the beta changing under you.

As of September 11, 2026 the Agents API costs nothing on top of tokens, tools and container time, runs in your own sandbox or one of nine partners', and still will not give you Zero Data Retention.

Frequently asked questions

Do you still need to host your own agent harness in 2026?

For most teams, no. OpenAI's Agents API entered public beta on September 10, 2026 and charges no fee for the harness itself, only tokens, tools and container time, and it will run the agent's compute in your own self-hosted sandbox. You still host your own if you need Zero Data Retention or non-US data residency, if you need hard egress control over the orchestration layer, or if your core loop cannot sit on one vendor's public beta.

What does the OpenAI Agents API actually cost?

There is no additional fee for the Agents API. You pay model tokens at the selected model's normal API rate, OpenAI's built-in tools at their standard rates, and container time for whatever sandbox runs the work. Anthropic's comparable product, Claude Managed Agents, has charged standard token rates plus $0.08 per session-hour since April 8, 2026, so the container line is the one to model before you migrate.

Does self-hosting the Agents API sandbox make it Zero Data Retention eligible?

No. OpenAI's own documentation states that the Agents API supports data residency only in the United States and does not support Zero Data Retention, and that choosing a self-hosted sandbox does not make the Agents API ZDR-eligible. Self-hosting moves where the agent's code executes. It does not move where the session state lives. That single clause is what keeps regulated workloads on their own harness.

What is programmatic tool calling and why does it lower token cost?

The model writes JavaScript that calls your tools, and that program runs in a fresh isolated V8 runtime with no network access, no Node.js, no package installation, no general-purpose filesystem and no subprocess execution. It can call tools in parallel, loop, branch, and filter or combine results in code, so only the relevant slice of a large tool result ever enters the context window. A 10,000-row query that gets filtered down to twelve rows inside the runtime is billed once at twelve rows, not once at 10,000 and again on every subsequent turn.

Which sandboxes does the OpenAI Agents API support?

Three options. An OpenAI-hosted sandbox billed at standard container rates, a self-hosted sandbox you provide, or one of nine partner sandboxes with first-class integrations: Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop and Vercel. The self-hosted option is what makes this a real build-versus-buy decision rather than a full handover of the execution environment.

Is adopting a vendor's agent harness worse lock-in than using their model?

Yes, materially. Swapping a model means changing a model string and rerunning your evals. Swapping a harness means rewriting where your orchestration logic lives, because session state, context compaction, crash recovery and subagent coordination all move into the vendor's service and out of your codebase. Two vendors now sell this shape, OpenAI since September 10, 2026 and Anthropic since April 8, 2026, and the two APIs are not interchangeable.

Is it safe to put a production agent on a public beta API?

It depends on how much of your product the harness is. Both OpenAI's Agents API and Anthropic's Claude Managed Agents are in public beta, which means the surface can change under you. Putting a peripheral internal workflow on it is a cheap bet. Putting the execution loop your product is sold on behind a beta header is a different class of risk, and the mitigation is keeping your tool definitions, evals and prompts in your own repo so the harness stays the replaceable part.

How do you test whether to migrate off your own agent harness?

Take one real workload you already have eval scores for, run it on the managed harness with a self-hosted sandbox, and compare three numbers: pass rate on your existing eval set, total cost including container time, and p95 latency. Do not compare against a demo. If the managed harness does not beat your own loop on your own evals, the only thing you bought was somebody else's maintenance schedule.