Agent Failure
Recipes

OPEN SOURCE · TYPESCRIPT · V0.1

Small failures.
Useful tests.

Your agent’s tool call can fail after the work is done. Find the bug, understand the repair, and take the test into your own code.

Get running in your terminal
THE RECIPE
  1. 01
    Reproduce the failureA controlled fault. A real state change.
  2. 02
    Make one repairSame starting point. Different behavior.
  3. 03
    Keep the regression testA small example you can adapt.

No model calls. No API keys. No account.

Pick a failure. Follow the evidence.

03 EXECUTABLE RECIPES

Controlled fixtures, not model benchmarks. The browser, CLI and tests run the same recipe code. All tickets, reports and service calls here are simulated.

FROM BROWSER TO CODE

Make this your next test.

Clone the repository. Run all three recipes locally. Then swap a fixture for your own disposable test service.

git clone https://github.com/swapupg/agent-failure-recipes.git
cd agent-failure-recipes
npm ci
npm run check
npm run recipes

Requires Node.js 24 or later. No runtime dependencies. Expected failures are verified by the teaching suite.

How to adapt a recipe ↗

A SMALLER STARTING POINT

Keep your stack.
Add an assertion.

These recipes teach what to check when the software around an agent fails. They fit alongside your existing tests.

Does a passing recipe prove my agent is safe?

No. It proves one invariant under one controlled fault. Your integrations, permissions, persistence, concurrency and model behavior need their own tests.

Why is the broken test suite green?

The teaching tests assert that the bug is reproduced. For an intentionally failing command, run a recipe with --variant broken. It exits with code 1.

What information leaves my browser?

The lab makes no model or tool API calls and sends no analytics. Theme preference is saved locally when available. GitHub Pages serves the files and may keep hosting logs. Links to external sources navigate to those sites.

Explore the concepts in Agent Explainer ↗