Guides
•3 min read
Customer Support Bot
Build a production support agent with PII redaction, auto-escalation, and budget controls.
The Problem
You need a customer support agent that handles real user data safely. It must redact personal information before processing, automatically escalate when it gets stuck, and not blow through your API budget on a single conversation.
The Solution
Combine createPIIGuardrail for data protection, constraints for escalation logic, and resolvers for the escalation action:
import {
createAgentOrchestrator,
createPIIGuardrail,
} from '@directive-run/ai';
const orchestrator = createAgentOrchestrator({
runner, // See Running Agents (/ai/running-agents) for setup
autoApproveToolCalls: true,
maxTokenBudget: 20000,
guardrails: {
input: [
createPIIGuardrail({
redact: true,
redactReplacement: '[REDACTED]',
}),
],
},
constraints: {
escalate: {
when: (facts) => facts.agent.turnCount > 5,
require: { type: 'ESCALATE', reason: 'Too many turns' },
},
},
resolvers: {
escalate: {
requirement: 'ESCALATE',
resolve: async (req, context) => {
// Your notification function — Directive doesn't provide this
await notifyHumanAgent(context.facts.agent.conversationId, req.reason);
context.facts.agent.status = 'paused';
},
},
},
});
How It Works
createPIIGuardrailscans input for SSNs, credit cards, emails, and phone numbers. Withredact: true, it replaces matches with[REDACTED]and lets the request continue. Withoutredact, it blocks the request entirely.- Constraints watch facts on every turn. When
turnCount > 5, the constraint fires and emits anESCALATErequirement. - The resolver handles escalation by notifying a human agent and pausing the orchestrator. The conversation can resume when a human takes over.
maxTokenBudgetprevents a single conversation from consuming excessive tokens.
Full Example
A complete support bot with PII protection, custom patterns, tiered escalation, and budget warnings:
import {
createAgentOrchestrator,
createPIIGuardrail,
createModerationGuardrail,
} from '@directive-run/ai';
// Block abusive input
const moderationGuardrail = createModerationGuardrail({
checkFn: (text) => {
const abusivePatterns = [/threat/i, /harm/i, /kill/i];
return abusivePatterns.some((p) => p.test(text));
},
message: 'I cannot respond to abusive messages. Connecting you with a human agent.',
});
// Redact PII with custom patterns
const piiGuardrail = createPIIGuardrail({
redact: true,
redactReplacement: '[REDACTED]',
patterns: [
/\b\d{3}-\d{2}-\d{4}\b/, // SSN
/\b\d{4}[\s-]?\d{4}[\s-]?\d{4}[\s-]?\d{4}\b/, // Credit card
/\b[A-Z0-9._%+-]+@[A-Z0-9.-]+\.[A-Z]{2,}\b/i, // Email
/\b\d{3}[-.]?\d{3}[-.]?\d{4}\b/, // Phone
],
});
const orchestrator = createAgentOrchestrator({
runner, // See Running Agents (/ai/running-agents) for setup
autoApproveToolCalls: true,
maxTokenBudget: 30000,
budgetWarningThreshold: 0.8,
onBudgetWarning: ({ percentage }) => {
console.warn(`Support bot budget at ${Math.round(percentage * 100)}%`);
},
guardrails: {
input: [moderationGuardrail, piiGuardrail],
},
constraints: {
softEscalate: {
priority: 10,
when: (facts) => facts.agent.turnCount > 3,
require: { type: 'SOFT_ESCALATE' },
},
hardEscalate: {
priority: 100,
when: (facts) => facts.agent.turnCount > 6,
require: { type: 'HARD_ESCALATE', reason: 'Conversation stuck' },
},
},
resolvers: {
softEscalate: {
requirement: 'SOFT_ESCALATE',
resolve: async (req, context) => {
// Inject a system hint to wrap up
context.facts.agent.systemPrompt += '\nPlease try to resolve this quickly.';
},
},
hardEscalate: {
requirement: 'HARD_ESCALATE',
resolve: async (req, context) => {
// Your notification function — Directive doesn't provide this
await notifyHumanAgent(context.facts.agent.conversationId, req.reason);
context.facts.agent.status = 'paused';
},
},
},
});
Related
- Guardrails reference — PII detection, moderation, and custom guardrails
- Orchestrator reference — constraints, resolvers, and lifecycle
- Prevent Off-Topic Responses guide — topic boundary guardrails
- Control AI Costs guide — detailed budget management
