The mandate
One or two sentences describing what an agent owns. It does not expire when you close the tab, and it is what the agent is measured against at the end of the month.
HOW IT WORKS
Staff, not software.
An assistant answers you. An agent finishes the task. The difference is six things — a machine of its own, a standing mandate, a hard approval gate, memory that outlives the chat, routines it writes itself, and colleagues it can hand work to.
01 — THE MODEL
A mandate, a machine,
and a line you draw.
The connections
Scoped access to the tools you already pay for — your books, your email, your repo, and a read-only bank feed through a licensed open-banking provider. No passwords, no browser automation, nothing screen-scraped.
The line
You set a threshold — an amount, a counterparty, a category. Below it the agent acts. Above it the work stops in your approvals with the reasoning attached. Filing, signing and paying are always above the line.
02 — A WEEK IN
What actually happens
after you sign up.
- 01
Day one · connect
You grant scoped access to your books, your email and a read-only bank feed. Zuger never sees a password, banking is read-only through a licensed open-banking provider, and every connection is revocable from one screen.
- 02
Day one · brief
You write two sentences per agent. Most founders start with Säntis and Titlis, because month-end and deadlines are what wake them at 04:00.
- 03
Day two · watch
The first run is deliberately loud. Every page opened, every field typed, every number and where it came from. You correct what is wrong; the correction is written to company memory.
- 04
Day three · shrink the noise
You raise the threshold on the things you no longer want to see. Most people stop approving invoice chases within a week, and never stop approving payments.
- 05
Week one · routines
Say "do this every month" and the agent writes the routine and spells the schedule back. Month-end close is the first one almost everybody keeps.
- 06
Week two · the second agent
Once one agent has your trust, the second is a five-minute decision. Dom usually arrives when the other agents start disagreeing.
03 — THE APPROVAL GATE
The part that makes
autonomy survivable.
These stops are enforced by the platform, not by the model. An agent cannot raise its own limit, disable a gate, or approve on your behalf, and the gate is the same whether the agent is running interactively or at 04:00.
| Action | Default | Configurable |
|---|---|---|
| Money leaving any account | Always stops | Threshold per agent, per counterparty |
| Filing with a registry or tax authority | Always stops | No — signature is always yours |
| Signing or countersigning a contract | Always stops | No |
| Anything reaching a customer or candidate | Always stops | Per channel, per agent |
| Changing a mandate or a threshold | Always stops | No — agents cannot edit their own limits |
| Reading, reconciling, drafting, scheduling | Runs freely | Can be tightened to stop |
| Internal handoff between agents | Runs freely | Can be set to notify |
04 — MEMORY
Corrections stick,
and they spread.
Company memory is not chat history. It is a set of decisions and preferences with a source, an author and a date, shared between agents and editable by you.
It records the decision, not the sentence
"Acme only signs annual, Dana approves, never quote monthly" — with a link to the thread where you said so.
It is shared, not siloed
The agent that learns your investor prefers Tuesdays passes it to the one booking the update. No agent has to be told twice.
You can read and delete it
Every entry is visible, attributable and removable. Deleting one takes it away from every agent at once.
05 — WHEN IT GOES WRONG
You find out from the log,
not from the consequence.
Everything is on the record
Every page opened, field typed and file touched, in an append-only log, with the reasoning that led to it. Exportable for your auditor in one click.
Most mistakes are caught at the gate
The consequential actions were queued for you anyway. The failure mode is a bad draft in your approvals, not a bad wire.
Revoking is instant and clean
Pull a credential and the agent loses access mid-task, without touching the other agents or the work already done.
QUESTIONS
The ones founders
actually ask.
What is Bernina?
Bernina is Zuger's guardian layer, named after Piz Bernina in Graubünden. It isn't an agent you hire; it inspects every Run the agents make. It's on by default and can't be switched off.
Does Bernina read my data?
Bernina inspects documents and outputs inside the same isolated Run as the agent and stores only a verdict — pass, redacted, blocked or escalated — plus references to the evidence. Redacted values are never shown in the app.
Do I need a technical co-founder to run this?
No. You need a bank account, whatever tools you already use, and ten minutes to say what each agent owns. The API exists for teams who want to build on top of it, not as a requirement.
So it does not actually file my VAT?
No — and that is deliberate. It reconciles the period, prepares the return in Xero and leaves it unsigned in your approvals with the workings attached. You file it. Filing to a tax authority needs recognised software and your own credentials, and we will not ask you for those.
What if an agent gets stuck?
It asks. A blocked agent posts what it tried, what it saw and what it needs — a login, a decision, a missing document — and then waits rather than improvising.
Can two agents contradict each other?
They can, and Dom exists to resolve it. When it cannot, the conflict is escalated to you as one decision, not two competing messages.
Does it work with a bookkeeper or lawyer I already use?
Yes, and most founders keep them. The agents prepare, evidence and chase; your advisers review and advise. The bill usually goes down because the preparation is already done.
Free plan · No card
See it run on your own books.
Connect one account, brief one agent, and watch a real close. Ten minutes, no card.