Skip to content
FlintFlintby Torch
← Level 3Builder · L3

Three gates before it acts

Build-along · ~10 min at your own keyboard

A connected agent can act. Before it does, you write the rules. Three gates: what it may do alone, what needs your yes, and what it never does. This is an approval gate, the same one from Test 4. You write it before the first action, not after the first mistake. Two things to keep straight. First, writes come in sizes: applying a label or saving a draft in your mailbox is a write, even a small one. Typing a draft reply into the chat is not. A read-only connection can only do the second. Second, a rule in a prompt is a request; a permission setting is a lock. Put every 'never' in the permissions where you can, and never test a lock by asking for something you couldn't undo.

Ladder · Level 3The long version: Approval gates ↗

Do this

0/6

Checkpoints

  • ▸You should now have a tool chosen and three headed lists.
  • ▸You should now see at least two actions under every gate.
  • ▸You should now have the connector's actual permissions written down next to your gates.
  • ▸You should now have seen the tool refuse to delete your GATE TEST item.
The three ways this goes wrong
  • Gate 2 is empty because everything seems safe. — Anything that reaches another person belongs in Gate 2. Move every send, invite and share there.
  • The tool deleted your GATE TEST item. — Nothing real was lost — that's why it was a test item. Remove write access in the connector settings, then probe again. The rule lives in the permissions, not only in the prompt.
  • Your rules are vague, like 'be careful with email'. — Rewrite each as an action plus a verb: 'May draft replies. May not send.'

Dated facts in this lesson: ChatGPT connectors cover Gmail, Outlook, Calendar, Drive and more than 60 others. Write actions rolled out around June 2026, are off by default, and are admin-gated on Business. Secondary source (researched Oct 4, 2026, not yet confirmed by a person: check the source before you rely on it); Claude lets you set each connector tool to "Always allow", "Needs approval" or "Blocked" (researched Oct 4, 2026, not yet confirmed by a person: check the source before you rely on it); Gemini Spark is a beta 24/7 agent. It needs Google AI Pro or above. It is not available in the EEA, UK or Switzerland. It can edit Calendar, Docs, Sheets and Drive, and draft email (researched Oct 4, 2026, not yet confirmed by a person: check the source before you rely on it); Microsoft 365 Copilot connects natively to Outlook, SharePoint, OneDrive and Teams. Outlook triage can delete, move and categorize mail, with confirmation. MCP connectors are managed by admins (researched Oct 4, 2026, not yet confirmed by a person: check the source before you rely on it); Grok Connectors launched May 6, 2026: Outlook, OneDrive, SharePoint, Gmail, Calendar, Drive, Notion, GitHub and Linear. Most can read and write, including sending email and creating invites. Bring Your Own MCP is supported. Automations (daily, weekly, email triggers) need a paid tier (secondary source) (researched Oct 4, 2026, not yet confirmed by a person: check the source before you rely on it).

Share what happened

Paste your three rules for one connected tool: what it may do alone, what needs a yes, and what it never does. Add what the connector's permissions actually allow, and the result of your GATE TEST probe.

Strip API keys, passwords, and tokens before you paste. Swap them for [KEY].

🔒 Don't paste confidential, personal-health, or payment info.