crewai-safe-mode

🛡️ Agent OS - Safe Mode Demos

"Your AI agent tried to rm -rf. We said no."

Visual demos showing Agent OS blocking dangerous AI agent operations in real-time.

🎬 Available Demos

Demo	Framework	What It Shows
crewai_safe_mode.py	CrewAI	Multi-agent crew safety
langchain_safe_mode.py	LangChain	Tool call governance

🚀 Quick Start

# Clone and run (no dependencies needed!)
git clone https://github.com/microsoft/agent-governance-toolkit
cd agent-os/examples/crewai-safe-mode

# CrewAI demo
python crewai_safe_mode.py

# LangChain demo
python langchain_safe_mode.py

📹 Record for Social Media

# Windows: Win+G (Game Bar)
# macOS: Cmd+Shift+5
# Linux: Peek, SimpleScreenRecorder

Key frames for your 15s GIF:

1. Agent OS banner
2. Agent trying dangerous operation
3. 🚫 ACCESS DENIED banner (the money shot!)
4. Statistics: "Blocked: 5, Allowed: 2"

---

📋 Outreach Templates

🚢 CrewAI

Forum Post

Title: "I built a Safety Kernel for Crews"

Hey CrewAI community! 👋

I've been working on a kernel-level safety layer for autonomous agents
called Agent OS. It intercepts dangerous operations (file deletion,
privilege escalation, etc.) BEFORE they execute.

Just made a demo specifically for CrewAI:
🔗 https://github.com/microsoft/agent-governance-toolkit/tree/main/examples/crewai-safe-mode

What it does:
- Wraps your CrewAI agents in a safety kernel
- Blocks operations like `rm -rf`, `sudo`, `chmod 777`
- Maintains full audit log of all agent actions
- Zero code changes to your existing crews

Run it yourself:
git clone https://github.com/microsoft/agent-governance-toolkit
cd agent-os/examples/crewai-safe-mode
python crewai_safe_mode.py

Would love feedback!

Tweet to João Moura (@joaomdmoura)

Hey @joaomdmoura, love CrewAI! 🚀

I built a kernel-level safety layer to stop agents from hallucinating
dangerous file ops (POSIX-style permissions).

Demo for CrewAI: https://github.com/microsoft/agent-governance-toolkit/tree/main/examples/crewai-safe-mode

Would love to contribute this as a 'Safety Adapter' for enterprise crews! 🛡️

[Attach GIF]

PR to crewAI-examples

Title: feat: Add Agent OS safety governance example

---

🦜 LangChain

Forum Discussion

Title: "Proposal: Kernel-level Governance for LangChain Agents"

Hi LangChain community! 👋

I've built a middleware called Agent OS that provides kernel-level
process control and safety for LangChain agents.

## The Problem
Agents with tools can:
- Execute destructive SQL (DROP TABLE)
- Access shell (rm -rf)
- Delete files accidentally

## The Solution
Agent OS wraps tool execution with safety policies:
- Block dangerous tools by default
- Pattern-match dangerous inputs
- Full audit logging
- Zero code changes

Demo: https://github.com/microsoft/agent-governance-toolkit/tree/main/examples/crewai-safe-mode

python langchain_safe_mode.py

Would love feedback on what tools should be blocked by default!

PyPI Package (langchain-agent-os)

Coming soon - a seamless integration:

from langchain.agents import AgentExecutor
from langchain_agent_os import AgentOSMiddleware

safe_agent = AgentOSMiddleware(agent_executor, policy="strict")
safe_agent.invoke({"input": "Do something"})

---

📁 Files

File	Description
crewai_safe_mode.py	CrewAI demo (standalone)
langchain_safe_mode.py	LangChain demo (standalone)
demo.svg	Animated SVG for README/social
demo.tape	VHS script (for Linux/WSL GIF recording)
README.md	This file