Cases: why trace every action an agent takes

In Scalt, every execution of an agent opens a case. It gathers what the agent received, what it read, what it did, what it asked a human and what it produced. Nothing happens outside it. This design choice is at the heart of the product, and it deserves an explanation. It has consequences for your teams, your costs and your audits.

Why trace

When a person makes a mistake, you can ask them what they did and why. An agent remembers nothing: if its work is not traced as it happens, a mistake remains a mystery. And a mistake nobody understands destroys trust more surely than ten explained ones.

The trace is not one more technical log, read only by developers. It is what lets the person approving do so with full knowledge, lets the team correct an agent on concrete grounds, and lets the company answer when asked how a decision was prepared.

What a case contains

• The trigger: a message, a command, an event or a scheduled job.
• The timeline of steps, each with its cost and duration.
• The documents produced, approved or rejected, and the assigned reviewers.
• Decisions and approvals: who approved or refused, and when.
• The version of the agent that produced the execution.
• Sub-cases and related cases, when one piece of work leads to another.

Cases can be viewed as a kanban board or a list, with filters and groupings. Each keeps its full timeline, and can be exported.

Approvals, in the right place

Tracing is not enough if approvals happen elsewhere, by email, late. In Scalt, when a rule requires a human, a card lands in the Inbox, with a Slack notification if you wish. The Inbox gathers what is waiting for someone: to approve, failed, done.

An action subject to approval stops and waits. You approve or refuse in the app, and the execution resumes straight away, or stops there. Either way, the decision is recorded in the case, with who made it and when.

For the documents the agent produces, such as a reply to a customer, a report or a draft contract, reviewers can be assigned. Each approves or rejects the document, and their decision joins the case. For every document that goes out, you therefore know who reviewed it and in which version.

Costs and circuit breakers

The cost of AI should not be discovered on the invoice. Every execution step has its cost, visible in the case, and every application shows its result indicators, its runs and its cost breakdown. Scheduled jobs, which run with nobody in front of the screen, have a budget and a circuit breaker: if they overrun, they stop. You therefore know what a given use actually costs, not just what AI costs in general.

An example: a customer credit note

A customer reports a double charge. The ticket triggers the support agent, which opens a case. The agent reads the contract and both invoices, finds the duplicate, and prepares a credit note along with a reply to the customer. Issuing the credit note requires approval: the card lands in the accounting team’s Inbox, with the documents the agent read.

The accountant checks, approves, and the execution resumes: the credit note is issued, the reply is sent. Three months later, during an audit, the question “who decided this credit note, and on what basis?” is answered by opening the case. Nobody had to piece the story together from emails and memories.

Correcting an agent on concrete grounds

When an agent gets something wrong, the case shows at which step: a poorly chosen source, a misunderstood instruction, a tool that returned an error. You correct the instruction, the source or the approval rule, then publish a new version of the agent. Since every execution keeps the version that produced it, you compare the cases before and after, and roll back if the correction brings nothing.

What it changes for audits

When an auditor, a customer or a regulator asks how an agent was supervised, the answer is not a document written after the fact: it is the history of cases, with their approvals, exportable. For the uses the AI Act classifies as high-risk, human oversight and log retention are deployer obligations; cases are useful material to meet them, not automatic compliance. The same history also serves internally: understanding why an agent was changed, and what the change made different.

Frequently asked questions