Agent settings
Choose the agent's AI key, name the platform, tell the agent what your system looks like and how its replies should sound, and give it a machine of its own.
The Agent page holds what the agent runs on and what it should know before it starts: the AI key, the platform's name, notes about your system, how replies should sound, and whether it has a machine of its own.
Save at the bottom
Make your changes, then click Save at the bottom of the page. Saved appears next to the button when it's done. If you leave without saving, your changes are lost. The other settings pages save on their own, so saving here never changes them.
The label at the top right says Ready once a key is chosen and saved, and Choose an AI key until then.
AI key#
The agent needs a Claude key to work. Keys are added once for the whole organisation, under Organisation › Keys, and each platform picks one here.
- 1
If your organisation has no Claude key yet, click Add a key (it opens Organisation › Keys) and add one. See Keys.
- 2
Come back to Settings › Agent. Choose the key in Claude key.
- 3
Click Save.
| Field | What it does |
|---|---|
| Claude key | The Claude key from your organisation's key library that this platform's agent runs on, or None chosen. Each option shows the key's name, whether it's an API key or a subscription token, and its last few characters. Until a key is chosen, the agent can't investigate this platform's tickets. |
With your own runner
When the agent runs on your own runner, the Claude key is kept by you, so the AI key row doesn't offer your key library. Choose how it is kept instead: Scrambled for your runner (you type it here, and your browser scrambles it so only your runner can unscramble it) or In your store (you put it on your runner's side and give us only its name). Chatbots and the help chat run in our cloud: they still use a key from your key library. See Two ways to keep your keys.
If the key disappears
If someone removes the chosen key from the library, the page says The key chosen before is no longer in the library. Choose another one. Choose another key and save.
Name#
| Field | What it does |
|---|---|
| Platform name | The name your team sees in the top bar. The agent also uses it in its instructions, for example "You investigate tickets for …". Up to 200 characters. |
About your system#
This is the most useful thing you can give the agent. It reads these notes before every investigation, so it knows where to look and what your words mean. The more specific you are, the better the investigations. Write them in Notes for the agent, up to 50,000 characters. The count shows under the box.
Good notes cover:
- What each repository does, and where the important parts live.
- Which services run where: Kubernetes namespaces, and the log labels that identify each app.
- Your product's terms, especially words that mean something different to you than to everyone else.
- Who your requesters are, and what they can and can't see.
- Known quirks: things that often look like bugs but aren't.
For example:
- acme/shop-api (Go): the shop's API. Orders are in internal/orders, invoices in internal/invoice.
- acme/shop-web (React): the dashboard our merchants use.
- Namespaces: shop-prod (API and workers), web-prod (dashboard).
- Loki: label app=<service name>. Errors are logged with level=error.
- A "merchant" is a business that sells through us. A "buyer" is their customer.
Requesters are always merchants. Buyers never contact us.
- Refunds take 3–5 business days to show. That is the bank, not a bug.
Describe repositories in two places
Each repository also has its own short description under Code. Use that for a one-line summary. Use About your system for how everything fits together.
How replies sound#
Tell the agent how to write to the requester (the person who opened the ticket), in Reply style. It follows this whenever it drafts a reply. Up to 5,000 characters.
Keep it short and direct. For example:
Plain language, no jargon. Concise, professional, to the point. Short answers.
Use Australian English. Address the requester by first name.
Never promise a date for a fix.
Time zone#
The agent writes every time in this time zone: in its findings, its questions and its replies, with the zone's abbreviation (for example 8:24 pm AEDT). Usually choose where the client is.
| Field | What it does |
|---|---|
| Time zone | The platform's time zone. Not chosen means times in UTC. When it isn't chosen, the page offers your browser's time zone with Use it. |
The times the app shows you, such as when a ticket was opened, are always in your own browser's time.
Its own machine#
A separate computer for each ticket, where the agent can run checks of its own. For example, it can look up a domain's mail and DNS records, test a website or a mail server, or check a certificate. It's off until you switch it on.
| Switch | What it does |
|---|---|
| Give the agent its own machine | On: while the agent looks into a ticket, it can run commands on its machine. It never has the machine while it writes a code fix. |
| Let it reach the internet | On: the machine can reach public websites and servers, never your private network. Off: it can only work with what's on the machine. Only shows as available while the first switch is on. |
- The machine holds none of your keys. It can't reach your code, cluster, logs, databases, Microsoft 365 or Google Workspace. Those stay with the agent's own read-only tools.
- Every command it runs, and what came back, shows on the ticket under what the agent did.
- The machine is removed when the agent finishes.
- When the agent runs on your own runner, the machine runs there too, and your runner's own rules decide as well: whether it may start one, and whether it may reach the internet (off until you allow it there). See The agent's own machine on your runner.
Think before you allow the internet
A ticket is written by someone outside your company. A ticket written to trick the agent could get it to send what it has read, such as a list of users, to an outside address. Switch the internet on only for platforms where these checks are worth it.
Click Save after changing either switch.