Tool Permissions

Understand tool permission types — allow, deny, and escalate.

ToolPermission

type ToolPermission = 'allow' | 'deny' | 'escalate';

Permission	Behavior
`allow`	Tool executes normally
`deny`	Tool call is blocked, error returned to LLM
`escalate`	Run pauses for human approval

PolicyCheckResult

When the framework checks a tool call:

interface PolicyCheckResult {
  allowed: boolean;
  decision: ToolPermission | 'escalated';
  reason?: string;
}

Per-Run Call Limits

Limit how many times a tool can be called in a single run:

policy: {
  tools: [
    {
      pattern: 'send-email',
      permission: 'allow',
      maxCallsPerRun: 5,  // Max 5 emails per run
    },
  ],
},

When the limit is reached, subsequent calls are denied.

Checking Tool Permissions Directly

import { AutonomyPolicy } from '@ahzan-agentforge/core';

const policy = new AutonomyPolicy(policyConfig);

const result = policy.checkTool('delete-user');
// { allowed: false, decision: 'escalate', reason: 'Destructive action' }

policy.recordToolCall('send-email');  // Track call count

Next Steps

Escalation — handling escalation requests
Budget — cost-based governance