OpenAI's Agents API Puts the Agent Harness Up for Rent. Should Enterprises Buy?

OpenAI announced on September 10 that its Agents API is available in public beta to all developers, offering the same managed harness that runs its Codex coding agent. The pitch is simple: instead of building the plumbing that keeps an AI agent working over hours, a company can rent it. For enterprise leaders trying to move agents from demo to production, the launch sharpens a decision that has been building all year. Should the agent harness be something you buy from a model provider, or something you own?

What is OpenAI's Agents API?

According to OpenAI's developer documentation, the Agents API gives applications access to the Codex harness through an OpenAI-managed service. OpenAI handles sessions, orchestration, context compaction and recovery. The customer supplies the tools and chooses where the agent runs, either in an OpenAI-hosted sandbox or on its own infrastructure.

The documentation describes support for web search, custom tools, Model Context Protocol (MCP) connections and delegation to subagents, and also lists computer use and plugins among the API's features. Sessions keep state across turns.

OpenAI says there is no extra fee for the harness itself. Customers pay standard API rates for model use and built-in tools, plus container rates for OpenAI-hosted sandboxes, the documentation says. CellCog chief executive Nitish Garg, whose company sells a competing product, described the launch as the harness becoming something developers buy rather than build. He also noted that the customer quotes OpenAI published at launch were chosen by the vendor, with no independent testing yet.

Two limits stand out for regulated companies. Data residency is available only in the United States, and Zero Data Retention is not supported, even if the customer runs a self-hosted sandbox, the documentation states.

Why does the harness matter so much?

An agent that runs for hours needs surrounding machinery: memory, context management, retries, tool permissions and recovery when something breaks. That layer is where many enterprise agent projects stall.

Recent research shows how hard it is to get right. A paper posted to arXiv on September 4 tested whether AI coding agents could build a production-ready customer-service agent from the kind of messy records a real client provides. The best configuration, Anthropic's Claude Opus 5 running in Claude Code, succeeded on only about 23.9% of simulated evaluations, compared with 82.2% for an expert-built reference. The authors said the AI builders made shallow data queries, rarely talked to the client and shipped the first design that worked.

The same day OpenAI launched its API, Cursor released a beta of Projects, in which a coordinator agent directs other agents rather than writing code itself, AI Weekly reported, citing Cursor's blog. According to AI Weekly, Cursor blocks autonomous pushes to main branches, a reminder that orchestration and guardrails are now selling points in their own right.

What are the arguments for and against a managed harness?

The developer response split along familiar lines. In a Hacker News discussion of the launch, supporters compared a managed harness to the move from bare-metal servers to cloud: someone else handles patching, scaling and uptime, and the team focuses on its own tools and data. One commenter said enterprise clients often feel more comfortable with a frontier lab's compliance posture than with an assembled open stack.

Critics raised three concerns.

Lock-in. The core model APIs are similar across providers, so swapping models is relatively easy. A stateful harness that holds sessions, memory and orchestration logic is much harder to move. One commenter pointed to OpenAI's earlier retirement of its Assistants API as a reason to be wary of heavily stateful services.

Data control. Without Zero Data Retention, some companies will be unable to use the service for sensitive workloads. Others worried that managed tool calls can move data before a human intervenes.

Model choice. Several developers argued that cheaper or open models handle most agentic work well, and that a harness tied to one lab limits the ability to route tasks to the right model for the job.

Both sides have a point. A team without deep platform engineering experience will likely reach a working agent faster on a managed harness. Building one is real work, as Coder's engineers showed when they walked through their experiments with harnesses and subagents in a AI Connect webinar. A company with strict data rules or a multi-model strategy may find the convenience costly later.

How have enterprise practitioners framed this choice?

Speakers at AI Connect events have argued that control, more than capability, decides whether agents reach production. At The AI Enterprise Conference 2026, Hemakumar Velugula of Xnode recalled a CIO telling him, "I'm not worried about proving the technology, but I'm unable to prove I'm in control of it." His case for a control plane that sits between agents and enterprise data applies directly here: a managed harness controls how the agent works, but the enterprise still has to control what the agent can reach.

Rajesh Sinha, founder of Fulcrum Digital, offered a related warning at the same conference about relying on any single vendor's oversight tools. "Have your own control panel, which is vendor agnostic," he said, as part of a seven-layer architecture for production agents.

What should enterprise leaders do now?

Separate the harness decision from the governance decision. Even if a managed harness runs the agent, identity, permissions, audit logs and cost tracking should live in a layer the company owns and can point at more than one provider.

Check the data terms before the pilot. U.S.-only residency and the lack of Zero Data Retention will rule out some use cases. Legal and security teams should review them first.

Test against messy, real data. The arXiv results suggest the gap between a working demo and a production agent is large. Evaluate any harness on the company's own records and workflows, not on vendor examples.

Plan an exit path. Keep tool definitions, prompts and evaluation sets in a portable form, and ask how sessions and artifacts could be exported if the service changes.

The Agents API is still in beta, and its terms may change before general availability. The bigger signal is that model providers now want to own the layer where agents run, and enterprises must decide how much of it to hand over.

Frequently asked questions

What is OpenAI's Agents API?

A public beta service, launched September 10, 2026, that gives developers OpenAI's managed Codex harness for durable cloud agents, with OpenAI handling sessions, orchestration and recovery.

How much does the Agents API cost?

OpenAI charges no separate harness fee. Customers pay standard API rates for models and built-in tools, plus container rates for OpenAI-hosted sandboxes.

Does the Agents API support Zero Data Retention?

No. OpenAI's documentation says the service does not support Zero Data Retention and offers data residency only in the United States.

Filed under: Agentic AI, AI Governance & Trust, AI Strategy & ROI