OpenAI’s new Agents API and hosted sandboxes make one thing clearer: agents are no longer just chat windows with better instructions. They are becoming workers with a place to operate, files to touch, tools to use, and costs to manage.
That changes the business question. The first question is not “which model should we use?” It is “where does the agent work, what can it touch, and who owns the result?”
OpenAI announced the Agents API and hosted sandboxes on September 10, 2026. The company says developers can now build and run cloud agents with the Codex harness, while OpenAI handles orchestration, long-running sessions, and context management. Agents can run code, work with files, produce artifacts, install packages, and use skills and plugins inside a managed environment.
TL;DR
Agents need a job site, not just a prompt. OpenAI’s hosted sandboxes are another sign that agent adoption is moving from conversation into work environments. Business leaders should treat this as an operating design problem: define the workspace, permissions, budget, handoff, and review loop before agents start doing real work.
The important part is not the API
The developer headline is the Agents API. The business headline is the sandbox.
A sandbox is the working environment around the agent. It is where code runs, files are handled, packages are installed, and artifacts are produced. OpenAI’s announcement says teams can bring their own sandbox, connect a sandbox provider, or use OpenAI-hosted environments. It also names nine first-class integrations: Blaxel AI, Cloudflare Dev, Daytona, DigitalOcean, E2B, Modal, Oracle Cloud, Runloop AI, and Vercel.
That list matters because the market is not only competing on model quality anymore. It is competing on where agent work happens.
For a normal company, that sounds technical until you translate it into operations. If an agent is helping finance reconcile reports, where are the source files? If it is helping marketing produce campaign assets, where does it store drafts? If it is helping support triage tickets, what customer data can it read? If it is helping operations update records, can it actually change the system or only prepare a recommendation?
Those are not engineering details. They are management decisions.
A collaborator needs a workspace
Most companies still talk about AI as if the interaction begins and ends inside a chat box. Ask a question. Get an answer. Copy the answer somewhere else. Maybe paste in a spreadsheet. Maybe ask again.
That mental model breaks down with agents.
An agent that can run code, handle files, and produce artifacts is closer to a junior collaborator than a search box. It needs a workspace. It needs a defined assignment. It needs constraints. It needs someone to review the output before the work moves downstream.
This is where organizations either get leverage or create mess.
A human employee does not become useful because you hand them a laptop. They become useful because they understand the work, the systems, the approval path, and what a good result looks like. Agents are the same, just less forgiving in some places and faster in others.
If the workspace is messy, the agent inherits the mess. If the instructions are vague, the agent produces vague work. If nobody owns review, the handoff fails quietly.
Hosted does not mean owned
OpenAI’s announcement says there are no additional fees for using the Agents API itself, with teams paying for tokens and tools. In the community discussion, OpenAI also clarified that hosted sandboxes use standard container rates and model usage is billed separately.
That is a small detail with a big operational implication.
Agent work has a cost shape. It is not only a software subscription. There is model usage, tool usage, sandbox runtime, storage, integrations, retries, and human review time. If leaders do not define the work tightly, they should expect surprise costs and fuzzy accountability.
This is especially true when agents move from demos into recurring workflows. A one-off agent that analyzes ten files is a test. A daily agent that reviews every incoming contract, writes summaries, checks exceptions, and prepares handoffs is an operating expense.
That does not make it bad. It makes it real.
The companies that win here will not be the ones with the fanciest agent demo. They will be the ones that can answer boring questions early.
What does this agent do every day?
Where does it work?
What files can it access?
What can it create?
What can it change?
What happens when it is wrong?
Who reviews the output?
Where does the cost show up?
That is the difference between an agent experiment and an agent workflow.
What leaders should do this week
Do not start with a platform bake-off. Start with one workflow where an agent needs a real workspace.
Pick something concrete: preparing a weekly sales brief, reconciling vendor invoices, summarizing support escalations, reviewing documents for missing fields, producing first drafts of client follow-ups, or turning meeting notes into assigned tasks.
Then define five things before anyone builds:
- The workspace: where the agent does the work and where outputs live.
- The inputs: files, systems, records, messages, or data it can read.
- The authority: what it can draft, create, update, submit, or trigger.
- The review loop: who checks the work and what counts as acceptable.
- The cost boundary: when the workflow should stop, ask, or escalate.
This is not glamorous. Good implementation rarely is.
The agent shift is not only about smarter models. It is about moving AI from conversation into production environments where work actually happens. OpenAI’s hosted sandboxes are one more step in that direction.
Once agents have a place to work, leaders have a new responsibility: design the place well enough that the work can be trusted.
Research and structure: Mai. Direction and voice: John Lipe.
Sources: OpenAI Developer Community, “Introducing the Agents API and hosted sandboxes,” published September 10, 2026; The Verge, “Due to GPT-6 Astra demand, OpenAI has paused new subscriptions to its $200 ChatGPT Pro tier,” published September 11, 2026.