Inspiration
The inspiration for Veritas Agent came from a universal pain point for developers and system administrators: the constant, reactive firefighting of infrastructure issues. The tagline, "Stop reacting. Start healing," captures the core problem perfectly. We wanted to move beyond simple monitoring and alerting, which often creates noise and requires manual intervention, to an intelligent system that can understand, diagnose, and propose fixes. We envisioned an agent that handles the operational "noise," allowing teams to focus on strategic, high-impact work.
What it does
Veritas Agent is an autonomous, conversational AI agent that transforms infrastructure management. It connects to your servers and, through a natural language interface on platforms like Slack, Email, Discord, and Telegram, it can: 1.) Monitor: Provide real-time status reports on server health, CPU, memory, and load. 2.) Diagnose: Intelligently parse user requests to understand intent and diagnose issues based on logs and system status. 3.) Propose Remediation: Suggest actionable fixes, such as restarting containers or freeing up disk space. 4.) Act (with Approval): Execute these remediation actions, but only after receiving explicit human-in-the-loop approval, ensuring safety and preventing unintended consequences.
How we built it
We built Veritas Agent with a focus on extensibility and clear separation of concerns. The core components include:
1.) Caspian SDK: This is the backbone for our conversational interfaces, allowing us to easily connect to Slack, Email, Discord, and Telegram with minimal boilerplate.
2.) LLM Integration (Gemini/OpenAI): We leveraged a Large Language Model to act as the agent's "brain." It's responsible for parsing user intent and determining the appropriate actions from a set of defined functions.
3.) Infrastructure Abstraction Layer (infrastructure.py): This layer contains the core logic for interacting with the infrastructure. Initially, we built a robust mock infrastructure simulator to develop and test the agent's logic without needing live servers. The code is designed so this mock layer can be easily swapped for real tools like Docker, Kubernetes, or Prometheus.
4.) Python: The entire agent is written in Python, with agent.py serving as the main orchestrator and llm_agent.py handling the LLM interactions.
Challenges we ran into
A significant challenge was designing the human-in-the-loop workflow. We needed a robust system that could request a confirmation, wait for a response, and handle timeouts or ambiguous replies, all while maintaining a conversational flow. The mock infrastructure was crucial here, allowing us to test edge cases safely. We also spent time ensuring the email integration was seamless, which required a detailed understanding of Gmail's forwarding and OAuth setup to get the agent to read and respond to emails correctly.
Accomplishments that we're proud of
We are extremely proud of creating a fully functional, conversational AI agent for infrastructure management in a short period. Key accomplishments include:
- Multi-Platform Availability: The agent works across Slack, Email, Discord, and Telegram, making it accessible wherever the team operates.
- The "Human-in-the-Loop" Design: Implementing a safe and intuitive confirmation flow for all critical actions is a standout feature that builds trust and prevents "oops" moments.
- Extensible Architecture: The clean separation between the core agent logic, the LLM, and the infrastructure layer means that the project is not a dead-end prototype but a solid foundation for production-grade tools.
- A Working Demo: Creating a fully functional demo with a mock infrastructure that clearly showcases the agent's abilities to diagnose, propose, and act on problems.
What we learned
Through this project, we solidified our understanding of several key areas:
- Agentic Design Patterns: We learned how to effectively structure an AI agent to interpret commands, make decisions, and call external functions in a deterministic and safe manner.
- LLM Integration: We gained practical experience in integrating LLMs for a specific, non-chatbot task, learning about prompt engineering and parsing structured outputs.
- System Integration: The project highlighted the complexities and nuances of integrating with various communication platforms, each with its own SDK and limitations.
- The Importance of Safety: Building the human-in-the-loop feature reinforced how critical it is to design AI systems that augment human decision-making, especially in sensitive areas like infrastructure, rather than replace it entirely.
What's next for Veritas Agent
The future of Veritas Agent is about moving from a powerful prototype to a production-ready solution. Our next steps include:
- Real Infrastructure Integration: Replacing the mock infrastructure layer with real integrations for Docker, Kubernetes, Prometheus, and AWS. This will make it a fully operational tool.
- Enhanced Diagnostics: Improving the agent's diagnostic capabilities, allowing it to correlate events across multiple systems and suggest preventative actions.
- Granular Permission Controls: Developing a more sophisticated permission system to define which users or roles can approve which types of actions.
- Rich User Interface: Building a dedicated dashboard to provide a visual overview of the system state, agent interactions, and historical actions.
Log in or sign up for Devpost to join the conversation.