Home

Welcome to AgentWatch V2

AgentWatch is a high-performance, real-time observability and execution safety framework for autonomous AI agents.

We are currently in V2, which represents a fundamental shift in our architecture and engineering philosophy. We have abandoned the speculative, heavy "academic" approaches of V1 in favor of the Invariant Lattice Architecture—a brutally fast, deterministic safety engine.

The Brutalist Philosophy

AgentWatch is built on a "brutalist" engineering philosophy:

1. Zero Bloat: If a feature isn't in the critical path of our current pipeline, it is deleted. We practice strict YAGNI (You Aren't Gonna Need It).
2. Deterministic Safety over LLM Auditing: We do not use LLMs to monitor LLMs. It is slow, expensive, and non-deterministic. V2 enforces safety through static lattice rules.
3. Extreme Performance: Our target overhead for executing safety intercepts on an agent action is ~0.1ms.

Wiki Directory

🏗️ Architecture & Internals

• Architecture V2 (The Invariant Lattice)
• Why we purged V1 (Migration Guide)

🤝 Contributing

• The Contributor Rulebook - MUST READ before opening a PR.
• Project Roadmap

🛠️ Operations

• Troubleshooting & FAQ