OpenAI updated its Agents SDK this week, and the most telling detail is what the update separates: thinking from doing.
The new architecture divides agents into two distinct layers. The harness handles the control plane — the agent loop, model calls, tool routing, approval workflows, tracing, and run state. The sandbox handles execution — running code, manipulating files, installing packages, reading from mounted cloud storage. The harness is the brain. The sandbox is the hands.
This isn’t just clean software design. It’s a direct answer to the question enterprises keep asking: where does my data actually go when an AI agent touches it?
The split matters because it gives companies a real choice. Sensitive coordination and decision logic can stay inside trusted infrastructure — your own systems, your own controls. The execution layer runs in isolated, sandboxed environments from providers like Docker, E2B, Vercel, or Modal. An agent can process confidential documents without those documents ever leaving a controlled environment.
The sandbox itself is genuinely capable. It offers resumable state, so agents can pick up where they left off across multiple runs. It supports memory persistence, meaning agents carry learning from previous sessions into new ones. Large datasets mount from S3 or GCS directly — no copying files into context windows. For the kind of work enterprises actually want AI agents to do — long, multi-step tasks that span days and involve real data — this is the infrastructure that makes it plausible.
OpenAI calls these “long-horizon” agents. The phrase has become something of an industry shorthand for the work that matters most: not quick question-answering, but extended workflows that require judgment, iteration, and tolerance for failure across dozens of steps. The new SDK is clearly aimed at this category.
The update launches first in Python. TypeScript support follows later. Pricing is standard API rates, open to all customers.
One thing that stands out: OpenAI is explicitly positioning the harness as something you bring to your own infrastructure. Users “go build these long-horizon agents using our harness and with whatever infrastructure they have,” per the company’s framing. That’s a meaningful posture — less “run everything on OpenAI’s platform,” more “here’s the control layer, plug it into yours.”
The agent wars are now an infrastructure competition, and OpenAI just drew a clearer architectural line.
