Agents

Some work is not a conversation. It is a job that repeats: check this every Monday, look at every new supplier, summarise what came in overnight. An agent is that job, saved once and run by Lucy without you in the chat.

Runs on its own

On a schedule, when a system calls it, or when you press Run.

Every run is a conversation

Open any run to see what it read, what it did and what it produced, like any other answer.

Safe by design

Limits per run, its own sign-ins, and no action that needs your approval unless an admin allowed it.

Note

Agents are switched on by your admin. If you do not see Agents in the sidebar, ask them. See Agents for admins.

What agents are good for

Team	Agent	What it does
Finance	Weekly custody fee check	Every Monday at 08:00, compares the week’s fees with last quarter and flags changes
Compliance	New supplier intake	When the procurement system posts a new ICT supplier, checks it against the DORA register fields and lists what is missing
Compliance	DORA register reminder	On the 1st of the month, lists register entries due for review
Operations	Overnight digest	Every weekday at 07:30, summarises unread mail from key clients and drafts replies
Risk	Contract expiry watch	Weekly, reads the contracts folder and lists agreements expiring within 90 days
HR	Joiner checklist	When HR starts a joiner, prepares the document and account checklist and a welcome email draft

Create one

1. Open Agents in the sidebar and choose New agent.
2. Give it a name and one line saying what it is for.
3. Write the instructions: what Lucy should do in every run, step by step, as you would brief a colleague.
4. Choose who it runs as, and its tools and limits (below).
5. Choose Create agent, then add a trigger.

Who it runs as

Option	What it means
You	With your access, your memory and your allowance. Its runs are your conversations, and you can continue one yourself.
Itself (service agent)	Its own identity, for the organisation. Only admins create it and see its runs, and its runs cannot be continued.

Tools and limits

• Search the web and Run code and make files are off until you tick them. Only turn on what the job needs.
• Memory can be off, or let the agent use what Lucy knows about your work.
• Limits per run: the number of tool calls (30 by default) and the minutes (15 by default). A run that reaches a limit stops and says so.

Tip

Write instructions the way you write a process: numbered steps, one action each, and say what the final output should look like.

Choose when it runs

Schedule

Pick days and a time in a time zone you choose, for example weekdays at 08:00 (Europe/Luxembourg). A schedule can run as often as every 15 minutes. If Lucy was unavailable at the time, it runs once when it is back, not once per missed time.

Webhook

A webhook gives the agent a private web address. When another system sends a request to it, the agent runs, with the request’s content as input. Use it to start an agent from your procurement tool, your ticketing system or a form. The address is secret, so treat it like a password. You can rotate it. The agent can also send its result to another address when it finishes.

Run now

Run starts it immediately. With input lets you give it something to work on, such as a supplier name or a file.

Give it its own sign-ins

An agent never uses your mailbox or files. It has its own connections.

• Choose Sign in for the agent and sign in at the service. Use your own account or another one, such as a shared mailbox.
• Tick only the apps it needs, for example Mail without Calendar.
• You can revoke its access at any time without touching yours.
• An agent cannot ask for approval. A tool that would ask you first, such as sending mail, is refused to agents unless your admin explicitly allowed it. Agents are best at reading, checking and drafting.

Follow what it did

• Each run shows how it started (schedule, webhook, manual), when, how long it took, and why it failed if it did.
• Open a run to read it like any conversation: the steps, the files it made, the answer.
• Filter to failed runs to see what needs attention.
• Agent runs stay out of your normal chat list so a schedule does not flood it. Open the agent to see them.
• If a run is still going when the next is due, the schedule records a skipped run instead of piling up work.

Costs and limits

Runs use the allowance of whoever the agent runs as, and your organisation’s budget. An agent that reaches a limit stops, and the run says so. Every run is also capped by its own limits on tool calls and minutes.

Good practice

1. Start with one small job you already do every week, and read the first few runs line by line.
2. Keep instructions specific. Say what to read, what to produce and what to do when something is missing.
3. Prefer read and draft. Let the agent prepare, and a person sends.
4. Check the failed runs weekly. A failure usually means the instructions or limits need adjusting.
5. Name an owner. If the owner leaves the organisation, their agents stop. Service agents are the place for shared jobs.