Documentation › Writing AI Test Cases

Writing AI Test Cases

User  One of the two main ways to build a test in MaxTAF. The other is the recorder.

An AI test case is a test you write in plain English. When you run it, MaxTAF drives an AI agent that controls a real browser to carry out your steps — and, for Maximo, it can record everything into a reusable MXML case that replays with no AI cost.

What an AI test case is

Alongside MXML, Java, Python and JavaScript, AI is a first-class case type in MaxTAF. Instead of code, its script is a list of natural-language instructions. At run time MaxTAF sends those instructions to an AI agent that opens a real browser and performs them — clicking, typing, navigating and asserting — the way a person following your steps would.

This is ideal for:

• Getting started fast — describe a test in words, no locators or code required.
• Exploratory and one-off checks — the agent adapts to what it finds on screen.
• Authoring by demonstration — run the AI case once against Maximo and let it hand you a deterministic MXML case for your regression suite.

Creating an AI case

1. Go to Dev ▸ Workspace and, on the List tab, click Create.
2. Give it a unique name and choose the AI type.
3. Open the Script tab and write your steps in plain English (see below).
4. Save, then press Run and confirm. Follow the browser live in the run panel.

Writing good steps

The agent is capable but literal — clear steps produce reliable runs. The following practices come straight from MaxTAF's verified example scripts:

Two useful literal instructions the agent understands:

• Navigate to <url> — go to a page.
• Evaluate expression: <JavaScript> — run raw JavaScript in the page (this is how the recorder is driven headlessly — see Capturing MXML).

A worked example (Maximo)

This verified script logs into Maximo, creates an asset, asserts a field value, and captures the flow as an MXML case:

Navigate to https://your-maximo/maximo/oslc/graphite/manage-shell/index.html#/main

Start recording by executing: window.postMessage({ type: 'MXTAF_RECORDER_CONTROL', command: 'startWithClear' }, '*')

Login with username "marko" and password "markodelic12345".

Navigate to the Assets application.

Click the New Asset button to create a new asset.

Wait for the asset form to load.

Type a test value into the Asset Number field.

Execute the assertValue command from the MAIN FRAME (not from inside the iframe):
window.postMessage({ type: 'MXTAF_RECORDER_CONTROL', command: 'contextAction', action: 'assertValue', elementId: 'ASSETNUM@483' }, '*')

Stop the recording: window.postMessage({ type: 'MXTAF_RECORDER_CONTROL', command: 'stop' }, '*')

Create the case: window.postMessage({ type: 'MXTAF_RECORDER_CONTROL', command: 'createCase', name: 'AssertValueTest_' + Date.now() }, '*')

Report what actions were recorded and verify the assertValue action shows the correct value and elementId.

Parameters that control AI runs

Set these as project or case parameters (see the Parameter reference). Defaults are sensible; you rarely need all of them.

Your AI provider (Claude, Gemini, OpenAI or a custom endpoint) and its key are configured once per project under Project Settings ▸ AI Configuration — not per case. See the User Guide.

How an AI case runs

Behind the scenes, an AI run flows through a small pipeline:

MaxTAF engine
   │  POST to the AI hub (mx.playwright.server)
   ▼
Playwright hub  ──picks an idle node──▶  Playwright node
                                            │  spawns the AI agent
                                            ▼
                     AI agent  ──drives──▶  real Chrome browser (with live view)
                               ──tools──▶   report pass/fail, screenshots, logs

The agent must finish by reporting a result; MaxTAF turns that into the run's pass/fail. If the agent stops without reporting, or hits a critical error (for example a detected loop), the run is marked failed.

Capturing a reusable MXML case

The big payoff for Maximo testing: while the AI agent drives the browser, the MaxTAF recorder (loaded in the same browser) captures the flow and produces an MXML case. That MXML case:

• replays deterministically with zero AI tokens — perfect for a regression suite;
• uses MaxTAF's Maximo-specialised language, handling iframes and dirty-record prompts;
• is uploaded straight into the project you named in mx.recorder.project_name.

So the pattern is: author once with AI, run forever as MXML. The headless recorder commands (startWithClear, stop, createCase, contextAction) are documented in the Recorder Guide.

Reading the results

Open the run in Ops ▸ Runs and expand the details:

• Console output — this is where the AI agent's live log goes: each assistant message, tool use, tool result, screenshot and error. This is how you see what the AI actually did.
• Log — separate, non-AI execution info.
• Run Report — a rich HTML report if the agent produced one.
• If recording was on, the generated MXML case appears in the target project's Workspace.