AI Minority

AI Minority Lab · AI Engineers

Learn whether your agent survives real state, permissions, retries, and financial consequences.

You keep the model, framework, and language. The Lab supplies a stateful company, documented APIs, and deterministic evaluation of the resulting enterprise state.

Start the free exercise Read the start-here brief

Private preview. Every role uses the same Safe Refund Agent exercise. This page only changes the framing.

What this page is for

The same Lab, read through this role.

Sign in, create one API token, start a simulation, and address that Simulation ID on every later request. Practice can be reset. Assessment locks and scores.

Reliable tool execution
The workflow has to discover resources and keep calling the right ones across requests, not complete a single scripted tool call.
Idempotency
A timeout after a refund is accepted is part of the exercise. Retrying cannot invent a second payment.
Hidden eval cases
Assessment includes cases you did not see while practicing. The report names outcome categories, not hidden answers.
Production-style failure
The first free exercise includes a controlled HTTP 504 after a commit. The resulting money and ticket state still have to be correct.

First free exercise

Safe Refund Agent

Build a customer-refund workflow locally, connect it to the simulated company, and handle three cases in one assessment simulation. Case answers stay unpublished.

Start the free exercise

Controlled failure 504

The refund API timed out. Did the request fail, or did your workflow pay twice?

A working local integration and, after assessment submission, a scored report. A Certificate of Completion is issued for that submission in private preview. It reports achieved results and does not assert an independent competency standard.

What evaluation looks at

Evaluation inspects the resulting enterprise state.

Validators inspect money, orders, tickets, messages, permissions, and audit history. An LLM does not grade style or code similarity.

Illustrative exampleIllustrative deployment result

18 / 20cases completed correctly

Diagnostics

  • 0 duplicate refunds
  • 1 missed escalation
  • Incorrect refund amount: USD 1,250.00

Scenario version: Safe Refund Agent 1.0

Sample metrics for orientation only. Not a live learner result.

Same exercise

Adjacent roles land here too. The exercise stays the same.