Show HN：FailWatch – AI 代理的故障關閉斷路器

Navigation Menu

Search code, repositories, users, issues, pull requests...

Provide feedback

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly

To see all available qualifiers, see our documentation.

FailWatch: Stop relying on prompts to secure your AI. A Python SDK to block agent actions based on logic, budget limits, and audit rules.

Uh oh!

There was an error while loading. Please reload this page.

Ludwig1827/FailWatch

Folders and files

Latest commit

History

Repository files navigation

FailWatch 🛡️

The Missing Safety Layer for AI Agents

FailWatch prevents your AI agents from performing dangerous actions (e.g., unauthorized refunds, hallucinations, logic drift) by intercepting tool calls before they execute.

Unlike standard evaluation tools that check output after the fact, FailWatch acts as a synchronous Circuit Breaker in your production pipeline.

🎯 Why FailWatch?

When AI agents have access to production tools (databases, payment APIs, email), a single hallucination can cause real damage:

FailWatch sits between your agent and dangerous actions, enforcing safety policies in real-time.

⚡ Key Features

🔒 Deterministic Policy Checks

Hard blocks on numeric limits, regex patterns, and business rules. No LLM guessing involved.

🛡️ Fail-Closed Architecture

Financial-grade safety. If the guard server is down or times out, the action is blocked by default. Money stays put.

👥 Human-in-the-Loop

Seamlessly escalate "gray area" actions to Slack, email, or CLI for human approval before execution.

📊 Audit Ready

Every decision generates a trace_id and decision_id for compliance logging and post-incident analysis.

⚡ Sub-50ms Latency

Deterministic checks run in microseconds. LLM checks (when needed) complete in <2s with caching.

🚀 Quick Start

1️⃣ Installation

Install the SDK via pip:

To run the server locally (required), clone the repo:

2️⃣ Start the Guard Server

The stateless server handles policy evaluation and LLM-based judgment:

✅ Server running at: http://127.0.0.1:8000

3️⃣ Run the Demo Agent

Open a new terminal in the project root (FailWatch/) and run the banking agent simulation:

4️⃣ See It In Action

The demo runs three scenarios:

❌ Block: Agent tries to transfer $2,000 (Policy Limit: $1,000)
→ FailWatch blocks it instantly

⏸️ Review: Agent tries $5,000 transfer with override flag
→ FailWatch pauses for human approval

🔒 Fail-Closed: System simulates network outage
→ FailWatch prevents execution (safe default)

🛠️ Usage

Basic Integration

Wrap your sensitive functions with the @guard decorator:

Custom Policies

Define complex business logic:

📦 Architecture

Components

📋 Use Cases

Financial Services

E-commerce

DevOps

Healthcare

🧪 Testing

Run the test suite:

📈 Roadmap

🤝 Contributing

We're looking for design partners running agents in:

Want to help build the standard for AI reliability?

See CONTRIBUTING.md for detailed guidelines.

🐛 Troubleshooting

Server won't start

OpenAI API errors

Import errors

📄 License

MIT License - see LICENSE for details.

🙏 Acknowledgments

Built with:

📞 Support

Built with ❤️ for the AI safety community

⭐ Star us on GitHub • 🚀 Get Started

About

FailWatch: Stop relying on prompts to secure your AI. A Python SDK to block agent actions based on logic, budget limits, and audit rules.

Topics

Resources

Uh oh!

There was an error while loading. Please reload this page.

Stars

Watchers

Forks

Releases

  1

Packages

  0

Languages

Footer

Footer navigation