Dunetrace is designed to enhance runtime observability for AI agents by automatically detecting structural failures and alerting users promptly. Unlike traditional monitoring tools that highlight issues post-factum, Dunetrace focuses on real-time problems, ensuring timely insights before users encounter errors. This proactive approach minimizes user disruption and enhances operational efficiency.

Key Features

• Real-time Monitoring: Observes the structural pattern of every run, providing alerts within 15 seconds of completion if it detects anomalies.
• Comprehensive Detection: Automatically identifies and assesses 15 structural issues including tool loops, retry storms, context bloat, reasoning stalls, and prompt injection, among others. Each detected issue comes with a clear explanation and suggested remediation steps.
• Privacy by Design: Ensures that no raw content leaves the agent's process, employing SHA-256 hashing for security prior to transmission.
• Versatile Integrations: Compatible with various frameworks including LangChain, FastAPI, and Flask, allowing for seamless integration into existing systems.
• Engaging Dashboard: Provides an intuitive dashboard that auto-refreshes every 15 seconds, displaying key metrics and alerts in real-time. Access it at http://localhost:3000.
• Alert Mechanisms: Supports alerts through Slack and other webhook systems like PagerDuty, enabling prompt notifications for critical issues.

Usage Example

To monitor an AI agent, include the Dunetrace SDK and easily instrument the agent:

from dunetrace import Dunetrace

dt = Dunetrace()  # No API key needed for local development
dt.init(agent_id="my-agent")  # Patches relevant libraries globally

@dt.agent()
def run_agent(query: str) -> str:
    ...  # Automatically tracks LLM and HTTP calls

For further details on employing Dunetrace to its full potential, refer to the documentation.

Dunetrace empowers organizations to maintain operational integrity and enhance the reliability of their AI systems.