Guides
How to let an AI agent work on your board safely
Updated 5 min read By Operators team
To let an agent work on your board safely, give it the smallest reach that still lets it do its job: its own named key, only the companies it needs, read only at first, and a rule that anything needing a decision comes to you as an approval request. Everything it changes is logged under its name and can be undone.
What are the controls?
| Control | Where you set it | What it protects |
|---|---|---|
| Its own name and access | Assistants and agents, one key or connection per agent | You can see exactly what each one did |
| Per-company access | Ticked when you create or connect it, editable later | A client’s agent never sees another client |
| Read only | A switch on each agent | It can look but never change anything |
| Approval requests | The agent asks, you choose Approve or Send back | Decisions stay with you |
| Undo | Activity, filtered to assistants and agents | Mistakes are cheap to reverse |
How do assistants and agents connect?
Assistants you chat with, like Claude or ChatGPT, sign in as you through the assistant link. You tick their companies on the approval page. Their changes appear as “Claude (for your first name)”.
Agents that run on their own, like a nightly script or Claude Code on a schedule, use an access key. Create one on Assistants and agents under New access key:
- Give it a name. This is how its changes are labelled, for example “agent: invites”.
- Tick the companies it may see.
- Leave Read only ticked for now: a new key starts read only.
- Create the key and copy it straight away. It is shown once and cannot be shown again.
The page also shows a ready-made command for using the key with Claude Code. Keep the key out of shared files and chat threads.
Which companies should an agent see?
Nothing is ticked by default, and an agent with no companies sees nothing. If you work for several clients, give each client its own agent rather than one agent with every company ticked. When one engagement ends, you remove that agent and nothing else changes.
If an agent asks for something in a company it cannot see, it gets nothing back. The attempt is noted in your security events in Settings.
Start read only, then widen
It starts read only. Untick it when you trust it. Run a new agent read only for its first few runs, ask it to report what it found, and check that it is reading the right companies and the right items. Then untick Read only from Edit access on its row.
Changes to access apply the next time it uses your board.
How should an agent ask for a decision?
An agent with changes allowed can add items, update them, mark them done and add notes. For anything where a person should decide, it should flag the item for your approval instead of acting. A flagged item:
- Moves to the top of Waiting on you on Today.
- Shows the agent’s reason and, if it left one, a link to a draft.
- Offers Approve and Send back. Send back takes an optional note on what should change.
The agent reads your decision the next time it checks the board. A flag stays until you act on it, even if the agent never returns. You get an email when an assistant asks for approval, unless you turned it off in Settings under Email and notifications.
Write the rule into the agent’s own instructions. For example: “Draft invoices and add them as items. Flag every invoice over the usual amount for my approval. Never mark an approval done yourself.”
What does the roster show?
Each agent has a row on Assistants and agents showing:
- When it was last seen, and how many changes it made this week.
- Which companies it can see.
- What it asked you to approve, and the outcome.
- What it did recently, from the summaries it reports.
- A Quiet badge when it has missed its usual rhythm.
- A notice if it made far more requests than usual.
Check any agent that has gone quiet or become much busier.
How do you undo an agent’s change?
Open Activity and choose the Assistants and agents filter. Every change shows who made it, what changed and when. Choose Undo on any row to put the item back. Undo covers changes on your board; nothing an agent does reaches your tools. If someone has changed the item since, Operators tells you it changed rather than overwriting their work.
If two people or agents change the same item at once, the later change is refused with a message to read the item again. Nothing is silently overwritten.
How do you remove an agent?
Choose Remove on its row. It loses access at once, and its key stops working. What it did stays in Activity. If you delete a company, every agent’s access to that company goes with it.
A weekly five-minute check
- Is every agent on the roster one you still use?
- Does each still need every company it can see?
- Is any agent quiet that should be busy, or busy that should be quiet?
- Are any approval requests older than a few days?
What to do next
Steps
-
Create one key per agent
On Assistants and agents, create an access key with a clear name, such as invites, so its changes are easy to spot in Activity.
-
Tick only the companies it needs
Nothing is ticked by default. An agent with no companies sees nothing at all.
-
Start it read only
A new key starts read only. Keep it that way for the first runs and check that it reads the right things.
-
Allow changes and ask for flags
Untick Read only when you trust it, and tell the agent to flag anything that needs a decision for your approval.
-
Review and undo
Each week, check its row on the roster, then filter Activity to assistants and agents and undo anything wrong.
Questions people ask
- What is the difference between an assistant and an agent here?
- An assistant is one you chat with, like Claude or ChatGPT, and it signs in as you. An agent runs on its own, like a nightly script or Claude Code on a schedule, and uses a named access key you create.
- Can an agent reach the outside world through Operators?
- No. Your board is private, and Operators only reads from your tools. Whatever an agent changes on the board stays on the board.
- What happens if an access key leaks?
- Remove the agent on Assistants and agents. The key stops working at once, and Activity still shows everything it did, so you can undo what you need to.
- Can an agent approve its own request?
- No. Only a person can approve or send back a flagged item. The agent can read your decision afterwards.
Related
- Guides How to connect Claude or ChatGPT to your Operators board Connect Claude, Claude Code or ChatGPT to your Operators board with one link, choose which companies it may see, and decide if it may make changes. 4 min read
- Articles How to delegate to AI agents without losing control A practical way to hand real work to AI agents, with clear briefs, narrow access, checkpoints for decisions and a record of everything they changed. 5 min read
- Articles Approve, review or watch: human-in-the-loop at work Human-in-the-loop explained for operators. When a person should approve, review or watch automated work, and how to keep the approval step meaningful. 5 min read
- Articles A founder's guide to approvals, what should wait for you Which decisions should wait for a founder's approval and which should not, with a simple test, written limits and a way to hand the rest to your team. 5 min read