The assistant

Agent in the sidebar is an assistant that works your data: it reads records, runs lookups, and — with approval — creates and changes them.

It sees exactly what your roles allow. A question about records you cannot read gets no answer, and an action you cannot run cannot be run on your behalf.

Asking

Ask in plain language. It works out which records are involved:

• How many sales orders are there?
• Show me all products.
• What would this customer pay for 40 of these on the 3rd?
• Create a new customer.

Each step appears as a card naming the step it took and where it got to — running, finished, or waiting for approval. Open a card to see the values it sent and the result that came back, so an answer can be checked against the steps that produced it. Answers are written out, with a table where a table is the honest shape.

When several records could be meant — two customers called Müller — it offers the candidates to pick from and carries on with the one chosen.

Approval

Reading never asks. Every change asks first.

Creating, updating, deleting, archiving, restoring, releasing a draft, and running any action all stop on the step’s own card and show the exact values about to be written. Two choices:

• Approve — it proceeds.
• Deny — it stops, and carries on the conversation without the change.

Nothing is written until Approve is chosen. A conversation left unanswered changes nothing. To send it down a different path, deny and say what to do instead.

Chats

Every chat is kept. The list on the left shows each one by its opening question, most recent first. Choose one to read it back and carry on where it stopped — from any browser signed in to the same account, so a chat started at a desk continues on a phone.

New chat starts a fresh thread. The assistant carries nothing across from the last one: what it knows about the work in hand is what the current chat says.

The bin icon beside a chat deletes it. The chat and everything said in it are gone and cannot be recovered; the records it changed stay, and stay visible in Record history.

Chats are private to the account that held them. Nobody else sees them, including an administrator.

Whoever runs OnERP sets how long a chat nobody has touched is kept before it is removed.

Agents that work on their own

Some agents are not asked; they start themselves when something happens. Order intake runs the moment a sales order is created: it reads the order, the customer, the stock of every product on it and the payment terms, and writes one note on the order with what it found — a line that cannot be covered from stock, terms that do not match the customer’s, or nothing off — and changes nothing else.

Such an agent acts as itself, under the roles an administrator has given it, not on anyone’s behalf. Nobody approves its steps: it may only do what its roles allow, and whatever the assistant would have stopped to ask about is refused instead, with the reason, and it reports that.

Every run is a chat the agent holds. It appears in the list on the left for every account, marked with a robot icon, and opens read-only: it can be read and deleted, but not continued. The records it changed show in Record history against the agent, with the chat as the cause.

What it cannot get around

The assistant uses the same rules as the screens. It cannot edit a released invoice, release a document missing a required field, or move stock that is held — those are refused the same way they are refused in a form, and it reports the refusal rather than working around it.

Changes it makes are recorded like any other, and show in Record history against the account that approved them.

Availability

The assistant is optional and has to be switched on by whoever runs OnERP — it needs a model provider configured on the server. Until then, opening Agent reports an error. See Run OnERP locally.