Rent the harness, keep the judgment

For two years, the hard part of running an AI agent was not the model. It was everything around it: keeping a session alive, trimming the context before it overflowed, running tools safely, handing work to sub-agents, and cleaning up afterward. That scaffolding is called a harness, and most small teams built their own or went without. If you are deciding who should build what, see Consultant, agency, build team or hire?.

On September 10, OpenAI put its own harness up for rent.

What shipped

The Agents API, in public beta, gives developers the same harness and infrastructure that run Codex. According to MarkTechPost’s write-up of the launch:

• Sessions are durable agent instances that hold a task and accept new input.
• Compaction is automatic. As a session nears its context limit, earlier context is condensed.
• Tools are searched rather than all loaded at once, and an agent can call several in parallel.
• Sub-agents take on pieces of a larger job.
• MCP servers, custom functions and built-in tools such as web search plug in.
• Sandboxes can be OpenAI’s own, your own machines through codex exec-server, or one of nine partners, including Cloudflare, Vercel, Modal and E2B.

There is no separate fee for the harness. You pay for tokens, tools and container time.

Two limits matter. Data stays in the United States only, and zero data retention is not supported yet. For regulated work, that rules it out for now.

What a business of one gets

The cost of running a real agent just dropped from “hire someone who can build a harness” to “write the job description well.” For an owner, that is the whole point. The scaffolding that used to take a team is now a line item.

What it does not give you

A harness runs an agent. It does not decide what the agent is for. Three things stay with the owner, and they are the parts that make an agent worth running:

1. The job. A clear statement of what done looks like, what the agent may touch, and what it must never do.
2. The knowledge. Your pricing, your clients, your voice, your past decisions. The harness can hold tools; it cannot invent what your business knows.
3. The stop. When the agent should stop and ask, and who checks the work before it leaves the building.

Renting the harness makes those three things more important, not less. When the machinery is cheap, the quality of the instructions becomes the difference between two companies.

Rent or run your own

A one-page way to decide:

• Rent the harness when the work is general (research, drafting, code in a repository), the data is not regulated, and you want to spend your time on the job description instead of the plumbing.
• Keep your own loop when the data must stay outside the US or must not be retained, when you need to switch models freely, or when the agent has to live inside systems a hosted sandbox cannot reach.
• Do both when you can. Most lean companies will end up with a rented harness for everyday work and a small loop of their own for the sensitive parts.

This is the kind of decision we make with owners in a consult block: which work moves to an agent, where it runs, and what the owner keeps. The tools will change again. The questions do not.