guardrails

Safety mechanisms that constrain agent behavior: input validation, output filtering, action limits, budget caps, and human-in-the-loop checkpoints.

Syntax

ai-agents

validate(input) -> agent -> validate(output) -> execute_if_safe(action)

Example

ai-agents

// Agent guardrails:
const safeAgent = {
  maxCost: 1.00,      // Stop if API cost > $1
  maxSteps: 20,       // Prevent infinite loops
  allowedTools: ["search", "read_file"], // Whitelist only
  bannedPatterns: [/rm -rf/, /DROP TABLE/], // Block dangerous actions
  requireApproval: (action) => action.type === "write_file",
  
  async run(task) {
    if (!this.isAllowed(task)) throw new Error("Task blocked");
    // ...
  }
};