ARC — Autonomous Reliability & Chaos Harness for AI Agents

ARC (Agent Reliability & Conformance Harness) is a deterministic simulation and chaos-engineering test harness for production AI Agents that perform real-world side effects (calling APIs, executing DB transactions, issuing payments, modifying cloud infrastructure).

🚀 Key Capabilities

• Zero-Prod Simulation & Fork-Replay: Replay production agent execution traces with deterministic virtual clocks and sandboxed tool models without touching production databases.
• Chaos Fault Injection: Inject systematic network timeouts, rate limits, schema mismatches, and double-action faults into agent decision loops.
• Automated Invariant Mining: Extract safety, authorization, and financial invariants from SQL DDL schemas, OpenAPI specs, and incident postmortems with human-in-the-loop approval.
• Delta-Debugging (ddmin): Automatically minimize long, noisy agent failure traces into the smallest reproducible YAML bug reproducer.
• CI/CD Quality Gate: Enforce statistical reliability verdicts in GitHub Actions or GitLab CI with standard POSIX exit codes (arc gate).

📦 Installation

pip install arc-reliability

For development and testing:

pip install "arc-reliability[dev]"

⚡ Quick Start

1. Instrument Your AI Agent with ARC SDK

from arc.sdk import trace_agent, trace_tool

@trace_tool(name="payment_gateway")
def process_refund(order_id: str, amount: float) -> dict:
    # Real API call or simulated response
    return {"status": "ok", "refund_id": "ref_123"}

@trace_agent(name="refund_agent")
def handle_customer_request(order_id: str) -> dict:
    return process_refund(order_id, 45.99)

if __name__ == "__main__":
    result = handle_customer_request("8842")
    print("Agent Execution Completed:", result)

2. Run ARC CLI Commands

ARC provides a unified command-line tool arc for debugging, verification, and invariant catalog management:

# Evaluate trace compliance in CI pipeline
arc gate --trace trace.json --budget 100 --alpha 0.05

# Replay and verify minimal bug reproducer in local sandbox
arc repro arc-repro-8842.yaml

# Start local replay debugger proxy
arc debug --port 8080

# Review proposed invariant candidates
arc catalog review --id INV-PAY-002 --accept

# Track and tag escaped incidents
arc incidents tag INC-1042 --candidate-id INV-PAY-002

🏗️ Architecture

ARC
├── src/arc/
│   ├── sdk/                 # Interceptor, Virtual Clock, Tier-1 PII Redactor
│   ├── cli/                 # Unified CLI application (`arc`)
│   ├── execution/           # Fork Engine, Chaos Injector, Sandbox Runner
│   └── control_plane/       # Approval Gateway, Trace Warehouse, Invariant Miner
└── schemas/                 # JSON Schema & SQL DDL definitions

🧪 Running Tests

# Run the complete test suite (Unit, Integration, E2E)
pytest tests/

# Run static quality & security verification
python3 pipeline/scripts/verify.py

📄 License

This project is licensed under the Apache License 2.0.