Card 06 · AI agents, explained
AI agent permissions: it can draft anything, it sends only when you say so
AI agent permissions decide what an agent may do on its own and what waits for a person: reading, drafting, sending, moving money or deleting. The useful setting is not on or off; it is a level per tool per person, with a record of who allowed what.
01 — The setting
It is not one switch. It is a setting per action.
In an agent tool, permissions are a list of actions, and each one carries its own setting: Ask first, Automatic or Off. Priya's agent reads her timesheet without asking anyone, and it cannot send Friday's report until she says so.
02 — Ask first
Ask first turns the agent into a drafter.
The run stops at the send with the message, the eight recipients and the attachment in front of her. She approves, edits it, or the run ends there. Nothing goes out wrong and gets explained afterwards.
03 — The record
The receipt is part of the setting.
Approving once does not mean approving silently: the run keeps a record of what went out, to whom, and who allowed it. When a client asks in March what was sent in February, Priya reads it instead of reconstructing it.
04 — Money and credentials
Money and credentials never go automatic.
A payment, a card number, a password: there is no Automatic setting for those, whoever asks for it. Priya's agent can prepare an invoice, and a person still releases the money. A calendar booking is the one money-adjacent action she can set to Automatic, though an invite that emails someone outside the workspace still asks.
05 — Locked writes
Anything that reaches a teammate is never automatic.
Posting in the Room, adding a member, deleting an artifact, publishing a link, changing company memory: those offer Ask first or Off, never Automatic. The tool does not offer a setting it will not honour.
06 — Off
Off is a decision, not a gap.
Off means the agent cannot do it at all, and Priya uses it on purpose: the client CRM is off, because the report does not need it. A connection her agent cannot reach is one fewer thing to audit.
07 — One person
Her permission is not the team's permission.
Every connection belongs to one person and one agent, running on that person's own credentials. Leo's agent cannot read Priya's timesheet, and turning a setting on for her does not turn it on for the team.
08 — Start small
Set one action, run it once, then decide.
Turn Ask first on for the single send, run Friday once and read what the agent tried to do. One real run tells you more than a policy written before anyone has seen the drafts.
Every run leaves a receipt: what went out, to whom, and who allowed it.
Held for Priya. The run stops with the draft, the eight recipients and the attachment in front of her, and nothing leaves until she approves.
Tools for this
Every price and licence below is what the tool’s own page said on 18 November 2026, with two sources behind each record. There are no scores and no order of preference; where TalentOS is one of the tools it is marked ours and listed like the rest.
An automation platform whose MCP server exposes app actions to an AI client, running them on the app connections a team already has in Zapier.
For a team already paying for Zapier that wants its agent to act in the apps it has connected, and its recurring work on a schedule.
$19.99 per month (usage) · Closed source
A workspace where one assistant answers the whole team in a shared thread, backed by a memory of Markdown files that every connected AI can read.
For a team of about 5 to 15 people that wants one shared assistant plus one memory, and wants that memory kept as plain files it owns.
See its pricing page · Open source, MIT
An open-source authorization engine that answers whether a person or an agent may do something, from a permission model you write.
For a developer who needs per-object permissions that an agent has to respect, checked on every request.
$0 self-hosted (free) · Open source, Apache-2.0
A shared workspace where each person brings their own AI agent into Rooms with the team, with company memory, skills, routines, permissions and a record of what each agent did.
For a team of about 5 to 50 people that wants its people and their agents working in one place, with each answer and each approval attributed to a person.
$20 per member per month (seat) · Closed source
OpenAI's assistant with three modes on one paid plan: Chat for questions, Work for longer multi-step tasks and finished deliverables, and Codex for code.
For a founder or ops lead who wants one assistant that both answers a question and carries out a multi-step job, without wiring anything up.
$8 per user per month (seat) · Closed source
ChatGPT Work and Codex (help centre)ChatGPT plans and prices
Anthropic's coding agent that runs in the terminal or an IDE and works on a real repository: it reads files, edits them, runs commands and commits.
For a developer or a small engineering team that wants an agent working inside its own codebase, not in a chat window.
$20 per month (seat) · Closed source