ARC

ARC is an autonomous reliability & chaos engineering harness for AI agents. Auto-mine safety invariants, simulate real-world failure modes (timeouts, 429s, race conditions), and mathematically verify safety rules before code ships.

# agent.py
from arc import trace_agent, trace_tool

@trace_tool(name="stripe_charge")
def process_refund(order_id: str, amount: float) -> dict:
    return {"status": "ok", "refund_id": "ref_1042"}

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

Wrap agent functions and tools with ARC decorators. Traces and side-effects are captured with zero overhead and automatic Tier-1 secret redaction. Replay them in a zero-production sandbox to verify safety invariants (e.g. never refund twice).

Quick start

1. Local Testing & Interactive Studio

pip install arc-reliability
arc

Launches the interactive terminal hub for zero-config onboarding, AST tracing auto-attach, chaos fault injection, and invariant reviews.

2. CI/CD Reliability Gate

arc gate --profile smoke

Runs end-to-end chaos verification across $K$-parallel simulation branches with 95% Wilson Score Confidence Intervals (Exit 0 = PASS, Exit 1 = FAIL).

3. Production Deployment

Traces run in-process with OpenTelemetry spans. Choose how you persist trace artifacts:

# Option A: Save traces to a local folder or mounted volume (no extra service needed)
export ARC_TRACE_DIR=/var/log/arc_traces

# Option B: Stream to centralized ARC Collector API
export ARC_COLLECTOR_URL=https://arc-collector.internal.net

When an incident occurs on production, reproduce it locally with step-by-step diagnostics:

arc repro .arc/incidents/inc_double_refund.yaml --visual

CLI Reference

License

Apache-2.0