Skip to content
HomeNewsHow OpenAI Runs Hundreds of OpenClaw Agents Safely, and How to Do the Same

Subscribe to our newsletter

One short email when we publish. No spam, unsubscribe anytime.

How OpenAI Runs Hundreds of OpenClaw Agents Safely, and How to Do the Same

How OpenAI Runs Hundreds of OpenClaw Agents Safely, and How to Do the Same

By OneClickClaw EditorialNovember 14, 2026

OpenAI runs hundreds of OpenClaw agents, and in a post on the OpenAI developer blog it listed the five controls that keep them inside clear limits. The post says the company runs "hundreds of OpenClaw agents" and that "we now have over 500 agents that run proactively in the background doing meaningful work". These are persistent agents: they stay on all day and act without being asked each time.

Kevin Lin of OpenAI shared the post on X on the 12th of November 2026, writing that it holds "some of the best practices we use to manage fleets of claws at @OpenAI". Peter Steinberger, who created OpenClaw, shared it too: "My team wrote about how we run persistent agents inside OpenAI".

OneClickClaw read the OpenAI post and the OpenClaw documentation on the 14th of November 2026. Below are OpenAI's five habits, each followed by the matching setting on your own agent.

What you need

  • An OpenClaw agent, on your own computer or on a server.
  • Its Control UI: the web page where you chat with the agent and change its settings.

Habit 1: run the agent apart from the part that controls it

According to the OpenAI post, the company splits each agent into two parts. The trusted part handles the network, the agent's identity (its own login, separate from any person) and its sandbox (a fenced area where the agent's tools run), and it always behaves the way it was set up. The agent itself, and any code it writes, is treated as untrusted, because an agent can write new code on the fly.

To do this, OpenAI added a feature to the open source OpenClaw project that lets the Gateway run in a separate container from the agent. The Gateway is the part of OpenClaw that stays switched on and connects the agent to your chats and AI models. A container is a sealed box that keeps one program apart from everything else on the computer. At OpenAI the Gateway even runs in a separate group of machines.

On your own agent, the closest step is OpenClaw's sandbox. According to the OpenClaw sandboxing documentation, the sandbox can be switched on per agent. When it is, the Gateway stays on the main machine and only the agent's tools move into the sandbox.

Habit 2: fence what the agent can reach

The OpenClaw crab pointing at a wooden gate in a server room, under the words Fence what the agent can reach

The OpenAI post lists four fences:

  • Network. Agents may visit only the sites on an approved list, and some sites are blocked by name.
  • Files. Agents may write only inside their own working folders.
  • Processes. Each agent runs apart from the other programs on the machine.
  • Secrets. Passwords and keys pass through a "credential broker", a middleman service, "so secrets are not visible to the agent".

OpenAI also published the settings it uses. In plain words, one example agent may talk to GitHub and nothing else, and sites often used to smuggle data out, such as pastebin.com and webhook.site, are blocked by name. Another setting keeps the agent away from the folder that holds SSH keys, the files used to log in to other computers. OpenAI also removed the option that would lift all of these fences at once.

On your own agent, the simplest fence sits in the chat box. In the Control UI, open the "Execution permissions" menu and pick one of five choices:

  • Default: follows the agent's own settings, which means full access unless someone set limits.
  • Read Only: the agent can read files in its folder and cannot change anything or run commands.
  • Guarded: the agent works in its folder, and you approve anything beyond it.
  • Workspace: the agent works in its folder, and a second AI checks anything beyond it.
  • Full Access: no limits and no reviewer.

Newer versions of OpenClaw show the same menu as "Permissions", with "Guarded" named "Ask me" and "Workspace" named "Auto-review", according to the OpenClaw permission modes documentation. If your agent's work fits in one folder, Read Only, Guarded and Workspace keep it there unless you, or the reviewer, approve more.

For keys, the OpenClaw page on secrets and storage says AI provider keys, such as a Gemini or xAI key, are refused when they sit in settings files inside the agent's own working folder. They belong in the Gateway's own settings, outside the folder the agent works in.

Habit 3: let a second AI check each action

The OpenClaw crab holding a signed approval card, under the words A second AI checks each action

The OpenAI post says that asking a person to approve every step leads to "review fatigue": people end up clicking yes to everything. So OpenAI uses Auto-review, where a second agent watches the first and approves its actions. Human attention is kept for the approvals that matter most. The post says this is built into OpenClaw, and that OpenAI turns it on "on all agent actions".

On your own agent, this is the Workspace (Auto-review) choice of the same menu. According to the OpenClaw permission modes documentation, the reviewer can allow an action, refuse it, or ask you. After three refusals in a row, the next decision goes to you.

For the commands an agent runs, the OpenClaw blog post on Auto mode says the reviewer can use a separate AI model. The agent can then do everyday work on one model while a stronger model checks its commands.

Habit 4: give each agent its own identity

According to the OpenAI post, every OpenClaw agent at OpenAI has its own identity, tied to a ChatGPT Service Account, an account made for a program. Each agent also gets its own key for the AI model, separate from the key of the person it works for. OpenAI says this lets it see which actions came from the agent and, if needed, switch off the agent's access to the model.

The same idea works for one person or a small team. Give the agent its own model key and its own accounts, separate from yours, so you can see its usage and switch it off alone. For an agent a team shares, the OpenClaw page on secrets and storage recommends a dedicated machine, a dedicated user, a dedicated browser profile and dedicated accounts, kept apart from personal Apple and Google accounts.

Habit 5: keep a full record

The OpenClaw crab holding a closed logbook, under the words Keep a full record

The OpenAI post says the company records "every configuration change on the control plane" and "all turns from the agent", meaning every setting change and every step the agent takes. Because those records can hold private details, only "a select few individuals in the security team" can read them.

For a single owner, the record already lives on the machine. The OpenClaw page on secrets and storage says to treat everything in OpenClaw's own data folder as if it holds secrets or private data. OpenClaw also has a built-in security audit, which the OpenClaw security documentation recommends after any settings change, to flag anything that has drifted from safe defaults.

A backup belongs in the same habit. The OneClickClaw guide on backing up OpenClaw and checking the restore walks through it.

What OpenAI is building next

The OpenAI post ends with OpenClaw Enterprise, an open source control centre for agents that OpenAI announced at its developer day. The post says an agent launched through it with the standard Codex preset (a ready-made setup for Codex, OpenAI's coding agent) gets "many of the safeguards" above built in. OpenClaw Enterprise is still in a pilot phase, and OpenAI plans its first full release by the end of the year. The OneClickClaw explainer on OpenClaw Enterprise covers what it is.

All five habits assume an agent that stays switched on around the clock. OneClickClaw runs OpenClaw and Hermes Agent on private servers that stay on around the clock, and new OpenClaw servers show the "Execution permissions" menu in the chat box.

Quick answers

How many OpenClaw agents does OpenAI run? The OpenAI post says "hundreds of OpenClaw agents", and "over 500 agents that run proactively in the background doing meaningful work".

What is Auto-review? A second AI checks what the agent wants to do and allows it, stops it, or asks you. OpenAI switches it on for every action. In OpenClaw it is the Workspace (Auto-review) choice.

Can I copy OpenAI's settings exactly? The settings in the post are written for Codex and name OpenAI's own folders and sites. The post itself says to swap the allowed sites for the ones your agent needs.

Do I need OpenClaw Enterprise for this? The permission menu, the sandbox and the security audit are all part of OpenClaw. OpenAI says OpenClaw Enterprise bundles many of these safeguards, and it is still in a pilot phase.

Helpful documentation

Want your own AI agent running without the setup? We host it on a dedicated EU server. Free for 7 days, no credit card.

Start your free trial

How managed OpenClaw hosting works

Get notified when we publish new articles

No spam, unsubscribe anytime.

By subscribing, you agree to our Privacy Policy.