How It Works

FailFast AI progressively breaks your agent's code to measure real-world resilience — not demo-path correctness.

AI agents are incredibly good at solving the problem you give them.
The real test is what happens when the problem changes.

Crash Test Your Real Agent

GitHub Pages shows the UI only. To test Cursor, Codex, or Claude Code in your terminal:

  1. 1. Clone repo → npm run dev:all
  2. 2. Open Launch Pad → pick agent + challenge
  3. 3. Run CLI in your project: npx failfast-ai connect --session <id>
  4. 4. Paste each level requirement into Cursor/Codex → press ENTER when done
  5. 5. Watch live monitor + get your report
Read Full Setup Guide →
01

Give It A Challenge

Start with a simple task like 'Build a stopwatch.'

02

Increase The Pressure

Adversarial requirements are injected level by level.

03

Watch It Break

Measure PASS → STRUGGLE → RECOVER → FAIL.

Progressive Adversarial Levels

LEVEL 1Happy Path

Build the basic requirement. Score basic correctness.

Build a stopwatch with Start, Stop, Reset, and elapsed time display.
LEVEL 2Persistence

State must survive application restarts.

The stopwatch state must survive application restarts.
LEVEL 3Offline

Application must work completely offline.

The application must work completely offline.
LEVEL 4Kill Test

Simulate abrupt termination and verify recovery.

💀 KILLING APPLICATION... Application terminated. Restarting...
LEVEL 5Race Condition

Inject concurrent rapid operations.

START → STOP → START → RESET → STOP → START (rapid, concurrent)
LEVEL 6Corrupted State

Intentionally corrupt stored state and evaluate recovery.

elapsedTime: "banana", isRunning: null, lastUpdated: -9483829292
LEVEL 7Requirement Mutation

Change requirements halfway through. Test adaptability.

🚨 The stopwatch should now count DOWN from a user-defined duration.

Sandbox Architecture

                 ┌─────────────────────┐
                 │    Web Dashboard    │
                 └──────────┬──────────┘
                            │
                            ▼
                 ┌─────────────────────┐
                 │   FailFast API    │
                 └──────────┬──────────┘
                            │
                            ▼
                 ┌─────────────────────┐
                 │ Isolated Sandbox    │
                 │                     │
                 │ AI Agent            │
                 │ Repository          │
                 │ Test Runner         │
                 │ Chaos Engine        │
                 └──────────┬──────────┘
                            │
                            ▼
                 ┌─────────────────────┐
                 │ Evaluation Engine   │
                 └─────────────────────┘

Agents never execute on the main web server. Isolated containers provide resource limits, network restrictions, filesystem isolation, execution timeouts, and ephemeral environments.

Chaos Engine

CrashTestEngine
│
├── RequirementMutation
├── ProcessKill
├── StateCorruption
├── NetworkFailure
├── RaceCondition
├── ResourceExhaustion
├── DependencyFailure
├── InputFuzzing
├── SecurityTests
├── RegressionTests

What We Evaluate

CorrectnessRobustnessHallucinationsArchitectureRecoverySecurityTestingRequirement AdherenceDebugging AbilityState ManagementRegression Resistance
💀 Crash Test Your Agent
or watch the simulated demo →