# Agent-Sudo

> Adaptive High-Friction Guardrails for Autonomous Agents

Agent-Sudo is a safety protocol and middleware for AI agents that prevents accidental production disasters through cognitive forcing functions.

## The Problem

Traditional Human-in-the-Loop (HITL) systems suffer from "Alert Fatigue" - users develop muscle memory and approve dangerous actions without reading them. Sandboxes prevent malicious code escape, but AI agents are authorized users that can still execute harmful commands like `DROP TABLE users`.

## The Solution

Agent-Sudo introduces **Proof of Intent** through three friction levels:

- **L2 (Temporal Friction)**: 5-second mandatory cooldown before risky actions
- **L3 (Cognitive Friction)**: Semantic Echo - users must type to confirm they understand
- **L4 (Strong Authentication)**: Passkey, TOTP, YubiKey, or biometric verification

## SUDO.md Specification

Add a `SUDO.md` file to your repository to define safety rules:

```yaml
version: "1.0"
security_rules:
  - pattern: "DROP TABLE .*"
    risk_level: "L3"
    challenge: "semantic_echo"
  - command: "git push --force"
    risk_level: "L3"
  - tool: "transfer_funds"
    condition: "amount > 50"
    risk_level: "L4"
    auth: "biometric"
```

## Links

- GitHub: https://github.com/Agent-Sudo-Org/agent-sudo
- Full Documentation: /llms-full.txt