Your first check,
step by step.

Try the isolated simulation.

Open the public simulation, choose synthetic order data and a complete or missing-material scenario, then inspect the application. Use the preview model when offered, or the built-in workflow, and start a test. Review its live BPMN trace, event timeline, assertions and evidence exports. Add feedback, edit the workflow and rerun a separate version. Only the in-memory sandbox executes; real SAP and workstations stay disconnected. Temporary runs expire after 30 minutes.

Request your invitation.

Ask Ed for an invitation, early-user service agreement and demo. Explain what you want to test and whether you need hosted BYOM or Enterprise. Pricing is pending. The email action opens your own mail client; nothing is automatically sent.

After the agreement, Ed supplies a private hosted member credential. Enterprise administrators provide their installation’s operator credential. Use Sign in; keep the access credential private. It is separate from a model API key.

Connect your own model.

In Model connections, choose OpenAI Responses, Anthropic or an OpenAI-compatible API. Enter the provider’s base URL, your API key and a model ID available to that API account. Confirm that the provider may receive your objectives and bounded application observations, then save the connection.

The key is encrypted on the server and excluded from reports. Hosted endpoints require approved HTTPS hosts; ask Ed to enable a custom provider. Enterprise additionally supports local Ollama and approved relay endpoints. Provider usage charges belong to your provider account. There is no automatic model fallback.

Check that the model responds.

Select the saved connection in the workspace and click Connect. Wait for a readiness response. A successful model-list request alone does not prove the model can plan a check. If the response fails, check the endpoint, API key, model availability and timeout before continuing.

Pair your workstation.

Open Connect computer. Download the published Windows bridge and run it without administrator rights in the Windows session containing your application. Download a one-use pairing file from ARBT, open it in the bridge and confirm the ARBT site address. The pairing file expires after ten minutes.

Keep the bridge running in the notification area. Hosted single-user access supports one workstation: revoke it before pairing a replacement. Enterprise supports multiple workstations. SAP GUI and its scripting permissions are installed separately. An unpublished bridge build must be supplied by your installation’s operator; source-listener setup is included with the Enterprise package. On Linux, run the supplied python -m desktop.app bridge in your graphical session; pair it with the same downloaded file. Local Enterprise users can choose This computer directly without pairing.

Select and inspect the application.

You can start with only the objective when a model and execution workstation are available. The agent discovers applications from open tabs, desktop sessions and URLs or paths in your prompt. Optionally choose your execution computer and application type to guide it. For a web app, enter one URL or local IP/port, or choose Attach to browser and find tabs. If no attachable browser is found, choose Open an automation browser, sign in to your test website there and find its tab. For Windows, select a running application or an executable. For SAP, discover the active session and choose the approved system/client. For Linux/GNOME, discover an accessible desktop application. For Excel, choose an .xlsx path and worksheet; edits export a separate copy. Available choices follow your selected execution computer. Click Inspect application and review observed controls. Do not guess SAP field IDs or use production data in the preview.

State the expected outcome.

For example: “Inspect the practice order. Verify Delivery status is Complete and Items is 4. Capture the evidence.” Describe your testing goal. Exact values, application details and procedure examples are optional context: the agent discovers controls, infers expected behavior, adds relevant risk checks and verifies them with actual assertions. Starting a run authorizes automation for the selected target and objective. There is no execution-policy selection or separate input-approval checkbox. You are responsible for the behavior you request; the application’s target and action boundaries still apply.

Run the check and inspect evidence.

Start the test. Watch observations, permitted actions and deterministic assertions. Pause or stop when needed; a current atomic action may need to finish. A model’s proposed coverage is not an executed result. Read failed or blocked steps, then export the current test’s JSON, HTML or workflow evidence.

For a revised check, review the typed workflow and rerun a separate version. A successful intermediate assertion does not establish that the full objective completed.

Use the right edition’s tools.

Hosted BYOM includes current-test evidence and one workstation, without fleet monitoring, session tracking, recorded learning or historical session browsing. Enterprise adds monitoring/history, multiple workstations and application-scoped interaction recording. Recording requires consent, stopping and reviewing captured steps before approval and replay.