How Faramesh Works

Action Flow

The complete flow from submission to result:

Agent → Submit Action → Policy Engine → Decision
                                    ↓
                            allow / deny / require_approval
                                    ↓
                            (if approval needed) → Human Review
                                    ↓
                            Execution → Result → Storage

Policy Engine

The policy engine evaluates actions using a first-match wins strategy with default deny:

• Policies are evaluated in order
• The first matching rule determines the decision
• If no rules match, the action is denied by default

Policy Evaluation Process

1. Action is submitted with: tool, operation, params, context
2. Policy engine checks each rule's match conditions
3. First matching rule determines: allow, deny, or require_approval
4. Decision is returned with reason and risk level

Policy Matching

Rules match actions based on:

• tool - The tool name (e.g., "shell", "http", "stripe")
• op or operation - The operation (e.g., "get", "post", "exec")
• pattern - Regex pattern matching against params
• amount_gt - Numeric comparison (for financial operations)
• context - Additional context like agent_id

Example Policy Match

rules:
  - match:
      tool: "shell"
      op: "*"
      pattern: "rm -rf"
    deny: true
    description: "Block destructive commands"
    risk: "high"

This rule matches any shell operation containing "rm -rf" and denies it.

rules:
  - match: { tool: "http", op: "GET" }
    allow: true
  
  - match: { tool: "*", op: "*" }
    deny: true
    description: "Default deny"