Skip to Content
Docs

Agents

What an agent is in Tosven, which agents exist today, who manages them and what they never do.

An agent in Tosven is a member of staff that does routine work and hands the decision to a person. It sits on the same roster as your people, with a written job, a human manager and an autonomy level you set. This page explains the roster, the role of the manager and the limits no setting can remove.

What an agent is

An agent is not a chat box. It is an account on your team roster with four things attached:

  • A job description. A written purpose that says what the agent is there to do.
  • A human manager. One named person is responsible for it. An agent without a manager cannot exist.
  • A permission boundary. The agent has its own grants, and each grant carries an expiry date.
  • An autonomy level. L0 to L3, set per agent and per type of action. See autonomy levels.

The interface always answers three questions about agent work: who did this, why, and how sure. Agents are named after their job, such as Ledger agent, and violet is reserved for them across the product.

The roster

Two agents are in production in the pilot. Four more are designed and on the way.

AgentStatusWhat it does
Ledger agentLive in the pilotProposes account, VAT code and period for purchase invoices, proposes bank matches, watches suspense balances and period close tasks
Collections agentLive in the pilotFollows overdue sales invoices and proposes payment reminders on a ladder of up to three steps
Sales agentOn the wayDaily follow-up cards from deal signals. Deal facts only, no scores for people
Treasury agentOn the wayProposes a payment run and prepares a payment file. A person uploads it to the bank
Purchasing agentOn the wayOrder proposals from lead times and minimum quantities, once stock and purchasing arrive
Compliance agentOn the wayWatches VAT deadlines, filings and period locks. It never files a return itself

The Collections agent is a rule engine. It makes no model calls, charges no fees and no interest, and stops at once on a pause, a dispute or a payment promise. A partial payment always goes to a person.

At Kaarna Group, Sara Nyström manages the Ledger agent and Elina Koskinen manages the Collections agent.

Managers and permissions

The manager is the ceiling. What an agent may do is the overlap of its own grants and the rights its manager holds at that moment. The overlap is worked out again for every decision and again when the work runs. It is never copied or cached.

  • If Sara loses the right to post purchase invoices, the Ledger agent loses it in the same instant.
  • Autonomy can only narrow this boundary. It never grants a right the manager does not have.

Roles and permissions for people are described in companies and roles.

How agent work reaches you

Agents do not change your records quietly. Work arrives as a decision card that states what the agent proposes, the reason, the records it rests on and how confident it is. A card that lacks the reason, the evidence or the confidence cannot be created.

You approve, edit or reject the card. Your approval is bound to the exact content you saw. If the data or your own permission changed before the work runs, nothing runs. See decisions and approvals.

When an agent does not know, it says so. It abstains with a reason, such as no history for this supplier or a source that is out of date, and tells you the next step. It does not guess.

Each run also has a budget for steps, time and cost. When the budget is spent, the agent stops and hands the work to a person. It does not retry.

What agents never do

Some limits are product rules. They do not appear in settings and no autonomy level lifts them.

  • No payments. No agent executes a payment. The planned Treasury agent will only propose.
  • No unsigned ledger entries. An entry reaches the ledger with a person's signature or under a posting rule that a person published.
  • No tax filing on its own. Filings stay under human approval.
  • No scoring of people. No agent rates the creditworthiness of a private person or the performance of an employee.
  • No full authority over a private person. Actions with a significant effect on a private person never reach L3 and always carry a way to object.
  • No disguise. An agent never presents its decision as a human decision and never produces a figure without evidence.
  • No deleting, sending or cancelling of custom records. For record types made in Builder, an agent can read and propose. Nothing more.

Stopping an agent

The kill switch is the control for stopping work at once. It has three scopes: a single run, one agent, or every agent in a company.

Pulling the switch takes one person and applies immediately. Releasing it needs two people. When Tosven pulls the switch itself, for example after a pinned model or rule version stops matching, the agent drops to L0. It can still propose, and it can no longer apply anything.

To hold an agent back for a while without stopping it, pause it in the run plan.

Next steps

  • Autonomy levels: what each level from L0 to L3 allows, and how a level is raised or lowered.
  • Decisions and approvals: how to read a decision card and what your signature covers.
  • Run plan: when agents work, and how to pause one or run it now.

On this page