# Add a function

> Gives an agent a fixed pipeline of steps it calls as one tool, for the part of its work that must come out the same every time.

**Skill for your coding agent:** [`mindset-function/SKILL.md`](https://docs.mindset.ai/skills/mindset-function/SKILL.md)

Building an agent goes in order: build the agent, connect your systems, add functions, scripts, Skills and widgets, test it, take it live, then put it in your product. Highlighted: Add functions, scripts, Skills and widgets.

## Ask your coding agent

```
Build a function that compares an invoice to its purchase order line by line and returns the differences, their total value and whether there are any. Test it on INV-4471 and on one invoice that matches, then give it to the invoice exceptions agent.
```

Skill: [mindset-function](https://docs.mindset.ai/skills/mindset-function/SKILL.md)

## What gets built

A [function](https://docs.mindset.ai/docs/functions): a fixed list of steps the agent calls as one tool. A step calls an operation, calls a model through an AI model [connection](https://docs.mindset.ai/docs/connection-types), works out a value, or chooses between values. Every step runs, and a function can't call another function. It's published as a version, and agents always use the published one.

## What you decide

- **Whether it's a function at all.** It gives the same answer for the same input. Comparing an invoice to its purchase order qualifies. Deciding whether a difference is acceptable under the contract is a judgment, and stays with the agent.
- **What it returns.** Return the value a [script](https://docs.mindset.ai/docs/write-a-script) decides on, such as whether there are any differences.
- **Its test parameters.** One set of inputs for each path, including one awkward case.
- **Whether it chooses by who the session is for.** To answer differently per plan, region or account, read the label your backend set as `attribution.<key>`, never an input a caller could send. Give each label value its own test case with `attribution` set, plus one with no labels at all.
- **Whether a rule must never be broken.** When the agent may write only within rules, such as never refunding more than was paid, the function checks the rules and writes only when none refuses, and the agent gets the function instead of the raw write. A rule in the agent's instructions is a request it can misjudge. Before such a function goes live, one test case where every rule holds runs once for real, against a test record.
- **Which failures to retry.** Retries are off unless you ask, up to 10 attempts in total.

A run an agent starts has a 300-second deadline and at most 1,000 outside calls. [Limits](https://docs.mindset.ai/docs/limits) has the rest.

> **Warning:** Every operation the function calls must be marked **Available to functions** on its connection, and tested, or publishing stops.

## Done when

- Its test parameters read **Verified**: every step ran.
- It's published, and an active agent version holds it. Publishing is refused while it calls an untested operation, unless an admin chooses **Go live untested**.
- The agent called it in a conversation and used what it returned.

## See it working

**Live demo:** [Ask your data](https://docs.mindset.ai/demos/ask-your-data). Ask sales questions and a dashboard draws the answers from what the agent’s functions return.

**Live demo:** [Store assistant](https://docs.mindset.ai/demos/store). The agent filters an outdoor store’s product grid, points at its pick and fills your basket.

## In your repo

Each function is a file in the agent's `functions/` folder, holding the exact input `create_resource` takes. Your coding agent publishes a change once its test parameters pass. See [Keep your agents in your code](https://docs.mindset.ai/docs/keep-your-agents-in-your-code).

## In the Studio

**Functions → New function** opens the **Function Builder**. Describe the function and it writes the steps, which you can't edit by hand. **Preview → Document** writes the function out as prose, and **Preview → Diagram** runs it on **Fixture input**. Press **Save changes**, publish it on the **Versions** tab, then add it on the agent's **Resources** tab.

![A function's Preview tab, with the fixture input above the diagram of its steps.](https://docs.mindset.ai/images/docs/studio/function-preview.png)
