Skip to content
The AI Agent Factory

Chapter 7 on today's products

Approved by
Panaversity ·
21 min read

This page backs 7.8. The lesson teaches the decision: which setting holds each line of your Authority Envelope, and which lines go on the gap list. This page holds the products' own settings, with their sources and the date each was checked.

Products change every few weeks. Check the date on each box before you rely on it, and open its source if the date is old.

LineAnthropic's ClaudeOpenAI's ChatGPT
The conversation settingManual or Auto. Skip all approvals in some placesOn the desktop, for local actions: mainly Ask for approval, Approve for me or Full access
Per-tool control (7.8's first difference)Per tool or group: Always allow, Needs approval or Blocked. Always ask, for one tool, is rolling outPer app: four levels. At Allow low-risk actions, the product decides what counts as low risk
Consequential actionsAsks before permanent deletion. Claude in Chrome is built not to take some actionsDesigned to ask before actions that are hard to undo, or that create a real-world commitment
Tool orderConnectors first, then the browser, then the screenA connected app first, then the cloud browser
Prompt injectionAn attack needs untrusted input and a harmful action. No browser agent is immuneAn attack needs a source and a sink. App safeguards do not remove the risk
Runs no one watches (7.8's second difference)Scheduled tasks: with a tool set to Always ask, the task may wait for you, or finish without the tool. Claude Code's scheduled tasks: runs start in auto modeA scheduled task may pause when an action needs approval
Admin policyOwners choose the connectors, and Enterprise roles can make tool permissions stricterAdmins choose each app's actions, and how new actions are treated

Figure 7.8. The same envelope on both AI vendors, from the boxes below, with their dates. Availability varies by product, plan and workspace. Documented behavior is no guarantee that an unsafe action will be blocked, so recheck your own settings before use.

The two AI vendors' boxes follow the same eight lines in the same order. The open-source box uses the lines its tools match, and it ends with how to run them safely.

Anthropic, as verified 18 November 2026.

  • The conversation setting. A permission setting in the message box covers the whole conversation.1 Manual pauses and asks you to approve actions, except for tools you set to Always allow.2 Auto keeps working, with automated safety checks before each action. Auto is rolling out as the default on Pro, Max, Ultra and Team. Manual is the default on Enterprise, except in the Cowork side panel in Chrome.1 Anthropic's other pages call these two Automatically approve and Manually approve.3 A third mode, Skip all approvals, does not pause to ask, and no safety check runs before Claude acts.4 Claude in Chrome offers it.3 Even in this mode, Claude asks before each use of a tool set to Always ask.2 In the one conversation, the message box offers only Auto and Manual.2 Anthropic's admin page says the third mode applies there to scheduled tasks and the built-in browser, and an organization setting can turn it off.5
  • Per-tool control (7.4). In the settings of each connector, a link from Claude to another app, you set each tool or group of tools to Always allow, Needs approval or Blocked.6 A fourth choice for a single tool, Always ask, is rolling out on paid plans. Claude then asks before each use of that tool, in every mode.2 On Team and Enterprise plans, organization owners can set the same choices for the whole organization. Read tools are grouped apart from write and delete tools, and members cannot change the owners' choices.7 On those plans, Anthropic's page says members approve connector tools that can write, in every task, unless an organization owner allows Always allow for them.8 Its guide to tasks says that in Auto, Claude decides for those tools.2 The two pages differ, so do not count on Auto to ask you first.
  • Consequential actions (7.2). Claude asks before permanently deleting files, in any mode.4 Anthropic says Claude in Chrome is built not to make purchases, create accounts or follow instructions found in emails or web pages. Anthropic also says it asks before deleting anything permanently, changing permissions or entering sensitive information.3 These statements describe how it is meant to behave. They are not a guarantee.
  • Tool order (7.5). Anthropic's computer-use page puts connectors first, as the fastest and most reliable path, then the browser, then the screen. Computer use is available on the Pro, Max and Ultra plans. On Team and Enterprise plans, it is in beta. It asks before each app and blocks investment and cryptocurrency apps by default. It has no sandbox, or closed-off space, between Claude and your apps.9
  • Prompt injection (7.6). Anthropic's safety page says an attack needs two things at once. Claude reads information from outside your trusted boundary, and Claude can take actions that could harm you.4 Its research says no browser agent is immune, or fully protected.10
  • Runs no one watches. You can choose an approval mode when you set up a scheduled task by hand. The help page says what happens in one case. If the task uses a tool set to Always ask, it may wait for your answer, or finish without the tool.11 Anthropic advises against scheduling tasks that send messages or are hard to undo.4 Claude Code has scheduled tasks of its own, a tool for developers, which it called routines until recently. A run starts in auto mode, can stop to wait for an answer, and acts as you.12
  • Admin policy (7.4). Owners choose which connectors members may use,7 and on Enterprise, custom roles can make tool permissions stricter, never looser.13
  • Standing worker, verified 18 November 2026. Claude Duet, in research preview, keeps working after you close its app. Before it sends a message, makes a payment, signs a document or deletes anything, it shows exactly what it will do. Then it waits for your approval.14 If you cannot approve at once, it finishes the rest of the task and brings the approval back to you later. Some actions are refused even with approval, such as trading securities or moving money between accounts.15

OpenAI, as verified 21 November 2026.

  • The conversation setting. In the desktop app, a permissions control sets local actions, the actions on your own computer. Its main choices are Ask for approval, Approve for me and Full access. Which ones you see depends on your app version and your organization. Approve for me sends a request to an automatic review instead of to you.16 OpenAI's prompting guide also advises adding a rule to your prompt. The rule requires your approval before ChatGPT sends, publishes or changes information other people rely on.17 That is advice about prompts, not a setting.
  • Per-tool control (7.4). Each app, OpenAI's name for a connector, can have one of four levels: Always ask, Allow read actions, Allow low-risk actions or Allow all actions. Allow read actions reads without asking and asks before any change. At Allow low-risk actions, the product decides which actions count as low risk.18
  • Consequential actions (7.2). ChatGPT is designed to ask you before actions that are hard to undo or that create a real-world commitment, such as a payment. When it works in its cloud browser, a browser on a separate computer in the cloud, it asks in the chat.19 These statements describe how it is meant to behave. They are not a guarantee. OpenAI lists sending messages, changing records, changing access, payments and sharing sensitive data as higher-risk actions.18
  • Tool order (7.5). OpenAI advises using a connected app instead of the cloud browser when the app can do the task directly. ChatGPT may also make that choice on its own.19
  • Prompt injection (7.6). OpenAI says an attack needs a source, a way to influence the system, and a sink, a capability that becomes dangerous in the wrong context. It says dangerous actions, or sending sensitive data, should not happen silently or without appropriate safeguards.20 Lockdown Mode limits outgoing requests.21 OpenAI says its app safeguards do not remove prompt-injection risk.22
  • Runs no one watches. A scheduled task may pause when one of its actions needs approval.23
  • Admin policy (7.4). Admins choose which read or write actions each app may use, and how new actions are treated.22 In workspace agents, write actions are set to Always ask by default.24
  • Standing worker, verified 21 November 2026. A dot, OpenAI's always-on agent, starts with built-in rules for when to act alone and when to ask. Its Custom Rules let you allow an action, require approval for it or block it. A check called auto-review looks at actions that could affect your accounts or share information. It checks them against your instructions, your Custom Rules and OpenAI's safety requirements.25 Custom Rules cannot turn off the core safety requirements. The most sensitive actions, such as changing a password or moving money, need you to take over and finish them yourself. Approving one message gives the dot no ongoing permission. An approval given in advance covers only what you authorized.26

Open source, as verified 23 October 2026.

  • Per-tool control (7.4). In OpenCode, a permission rule sets a tool to allow, ask or deny, and the last rule that matches wins. Every agent starts from a base rule that allows everything. Only reading sensitive environment files asks first.27 In Omnara, each tool is set to always allow, always ask or always deny. Tools reached through MCP (Model Context Protocol), a standard way to connect AI apps to other systems, start at always ask. Built-in and custom tools start at always allow.28
  • Prompt injection (7.6). OpenClaw's README says to treat incoming messages as untrusted input.29 By default, an unknown sender gets a pairing code, and OpenClaw ignores the sender until you approve.30 Its security guide adds that the content itself is a threat, not just the sender.31
  • Runs no one watches. OpenClaw's first setup creates a Proactive check, a scheduled job that runs every 30 minutes by default.32 Omnara can send its approval requests to Slack or Discord.33
  • Running it safely. OpenClaw's sandbox, a closed-off space for its tools, is off by default, so its tools run on the computer itself.34 Before the first run, switch the sandbox on and write an ask or deny rule for each limit in your envelope.

MCP does not change any of these settings. The standard does not define an approval step, so the app's settings still do the asking.

The MCP standard. Its version of 28 July 2026, as verified 23 October 2026, says a person should always be able to deny a tool call.35 A server can now also ask for more input before it goes on.36

Chapter 7 on today's products

Question 1 / 2

0 answered

Anthropic's computer-use page orders three ways to reach a system. Which does it put first, and why?

  1. Claude Cowork and chat are one Claude, Claude Help Center.↑↑2
  2. Get started with Claude Cowork, Claude Help Center.↑↑2↑3↑4↑5
  3. Claude in Chrome permissions guide, Claude Help Center.↑↑2↑3
  1. Use Claude Cowork safely, Claude Help Center.↑↑2↑3↑4
  2. Set up and manage the new unified Claude experience for your Enterprise organization, Claude Help Center.↑
  3. Get started with connectors, Claude documentation.↑
  4. Use connectors to extend Claude's capabilities, Claude Help Center.↑↑2
  5. Use Claude Cowork on Team and Enterprise plans, Claude Help Center.↑
  6. Let Claude use your computer in Cowork, Claude Help Center.↑
  7. Mitigating the risk of prompt injections in browser use, Anthropic.↑
  8. Schedule recurring tasks in Claude Cowork, Claude Help Center.↑
  9. Automate work with scheduled tasks, Claude Code Docs.↑
  10. Manage custom roles on Enterprise plans, Claude Help Center.↑
  11. Introducing Claude Duet, Anthropic, 21 October 2026.↑
  12. Get started with Claude Duet, Claude Help Center.↑
  13. Permissions, ChatGPT Learn, OpenAI.↑
  14. Prompting, ChatGPT Learn, OpenAI.↑
  15. Managing app permissions in ChatGPT, OpenAI Help Center.↑↑2
  16. Using cloud browser in ChatGPT, OpenAI Help Center.↑↑2
  17. Designing AI agents to resist prompt injection, OpenAI.↑
  18. Lockdown Mode, OpenAI Help Center.↑
  19. Admin controls, security, and compliance for plugins and apps, OpenAI Help Center.↑↑2
  20. Scheduled tasks in ChatGPT, OpenAI Help Center.↑
  21. ChatGPT Workspace Agents for Enterprise and Business, OpenAI Help Center.↑
  22. Introducing dots, OpenAI, 29 September 2026.↑
  23. Dots privacy, security and safety FAQs, OpenAI Help Center.↑
  24. Permissions, OpenCode v2 documentation.↑
  25. Tool permissions, Omnara documentation.↑
  26. OpenClaw, OpenClaw Foundation, README on GitHub.↑
  27. Access control, OpenClaw documentation.↑
  28. Prompt injection, OpenClaw documentation.↑
  29. Heartbeat, OpenClaw documentation.↑
  30. Integrations overview, Omnara documentation.↑
  31. Sandboxing: modes, scope and backend, OpenClaw documentation.↑
  32. Tools, MCP specification 2026-07-28, Model Context Protocol.↑
  33. Multi Round-Trip Requests, MCP specification 2026-07-28, Model Context Protocol.↑
Owner
Panaversity

On this page