Skip to content

4. Build your own agent ​

Partners embed payroll in their own software: an accounting system's assistant, an HR chat, a voice channel. You own the UI and the system prompt; Salaxy provides the tools and the skills.

The contract: skills, not the agent file ​

The repository contains a reference agent (agents/payroll-agent.md), but your agent does not need it. Everything that must hold in every implementation is in the skills themselves — including the guardrails:

  1. Never invent an amount — ask.
  2. Never guess a wage type.
  3. Never invent an income-type code.
  4. Never invent an identifier — a personal identity code, bank account number or reference number is asked for; a guessed one can belong to a real person.
  5. Money is EUR.

The skills also carry rules for other people's payroll data — never write the nature of an illness into a reply, never repeat a personal identity code in prose, and a person, not the agent, decides that pay is paid — and the country boundary: Finnish rules are never applied to another country's payroll. See Privacy & AI compliance for the wider picture.

If you write your own system prompt, keep it about your product: tone, who the users are, what your UI can show. Leave payroll rules to the skills.

Pass your integration context ​

Some answers depend on things your software knows and the user does not: which payment channel the payroll runs through, whether a paid calculation can still be corrected, whether a service fee is shown. Pass them in your system prompt:

yaml
integrationContext:
  partner:        <your product name>
  paymentChannel: nonCfa | cfa | manualSepa
  accountTiers:   two | three
  userLevel:      entrepreneur | professional
  • These are skill values, not API values. paymentChannel here is the kind of channel the skills tell apart: cfa for a customer funds account (Salaxy pays), manualSepa for payment with a SEPA payment file, nonCfa for any other channel. It is not a value of the REST API's PaymentChannel enum. See Payment channels.
  • Every field is optional. A missing field means the agent lists the alternatives — not that it assumes the most common one.
  • A given field narrows the answer: with paymentChannel known, the agent says what applies to this user.
  • Context never relaxes checks. It changes what the agent says applies, never what it verifies. userLevel: professional means less explanation, not fewer checks.

Your users are not payroll professionals ​

The first users are small-business owners in a chat. The skills are written for that: the agent asks the business question, not the field question — "does the employer pay the fuel?" rather than "which data.kind?" — and explains terms the first time it uses them. When the user is a professional, it explains less, and checks just as much.

Don't send users to click ​

The skills assume the agent is the interface. They do not tell the user to press buttons in a UI the agent cannot see. If an action is not possible through the tools — correcting a paid calculation, for example — the agent says so and names what has to happen. If your product does have a UI for it, add that to your system prompt.

Verify ​

The eval suite drives an agent with the skills against the live tools and checks the result deterministically. It runs with Salaxy's reference agent instructions — your system prompt is a different configuration. Run the same cases with your prompt before you ship; each case is a JSON file with a prompt and the expected rows or the expected refusal.

Going live ​

The demo servers calculate on sample data. Real payroll uses the production servers — https://payroll.salaxy.com/mcp and https://reporting.salaxy.com/mcp — with the user's own Salaxy account; see Connect a client. The tools are the same, so an agent built on the demo carries over: change the address and sign in with a real account.

Every page is also available as Markdown: add .md to its URL — a folder page such as white-label/ is white-label/index.md. Index for LLMs: llms.txt. Questions: Support.