coreason-sentinel 0.3.0

alerts you when things go wrong in production

CoReason Sentinel

Production Monitoring & Circuit Breaker Service for AI Agents

Coreason Sentinel acts as a centralized "Watchtower" for distributed AI agents. It ingests cognitive traces (OTEL spans) and business logs (Veritas events) to monitor agent health in real-time. If safety thresholds are breached (e.g., high hallucination rates, budget leaks, or statistical drift), Sentinel trips a global Circuit Breaker, preventing further risky operations.

Architecture

Sentinel is built as a FastAPI service backed by Redis for state persistence.

• Ingestor: Asynchronously processes telemetry from agents.
• Circuit Breaker: Manages the global health state (CLOSED, OPEN, HALF_OPEN) based on configured triggers.
• Drift Engine: Performs statistical analysis (Cosine Similarity, KL-Divergence) to detect input/output drift.
• Spot Checker: Probabilistically samples traffic for deep evaluation (Assay).

Key Features

• Centralized Circuit Breaker: Shared health state across all agent instances.
• Real-time Monitoring: Tracks latency, token usage, cost, and custom metrics.
• Drift Detection: Detects vector drift and output distribution shifts using SciPy and NumPy.
• Control API: Simple REST endpoints for agents to check their status before execution.

Getting Started

Refer to the documentation in docs/ for detailed instructions.

Installation

pip install coreason-sentinel

Quick Start

1. Start Redis.
2. Run the Sentinel service:

   uvicorn coreason_sentinel.main:app --reload
   

3. Check the health endpoint:

   curl http://localhost:8000/health
   

License

Prosperity Public License 3.0