IT Help Desk
Triage Agent

The Problem This Agent Solves

Help desk teams lose significant time manually sorting, categorizing, and prioritizing incoming IT support tickets. Without a structured triage system, urgent tickets can get buried, response times suffer, and technicians waste time on tickets that should have been escalated or routed differently from the start.

AGENT-001 was built to demonstrate how AI can be applied directly to this problem — automating the triage decision, drafting the first response, and flagging escalation needs so a technician can focus on resolution rather than classification.

Architecture — How the Agent Works

The agent uses a tool-use architecture built on the Claude API. When a ticket is submitted the agent does not simply generate free-form text — it calls structured tools that return specific, typed outputs for each triage decision.

• classify_ticket — determines the ticket category based on the issue described
• assign_priority — evaluates urgency based on impact, affected users, and business criticality
• match_knowledge_base — searches for relevant articles or known fixes that apply to the issue
• draft_response — generates a professional first-response message aligned with the classification
• flag_escalation — determines whether the ticket requires escalation and documents the reason

This tool-use pattern ensures the agent produces consistent, structured outputs rather than unpredictable free text — making it suitable for integration into a real help desk workflow.

Building the Ticket Dataset

The simulation was built around 22 synthetic IT support tickets covering the most common help desk scenarios encountered in real-world IT support environments:

• Password resets and account lockouts
• Hardware failures — laptop, monitor, keyboard, printer
• Software installation and application errors
• Network connectivity and VPN issues
• Email and calendar problems
• Access requests and permission issues

4 deliberate edge cases were included to test the agent's real-world judgment — tickets that are ambiguous, have conflicting signals, involve security implications, or require escalation that is not immediately obvious from the surface description.

Step-by-Step Build Process

• STEP 01Defined the core triage workflow — what decisions a help desk technician makes on every ticket and in what order
• STEP 02Designed the tool-use architecture — broke each triage decision into a separate Claude API tool call with typed input and output schemas
• STEP 03Built the ticket dataset — wrote 22 realistic IT support tickets across 5 categories with intentional variation in complexity and urgency
• STEP 04Designed 4 edge cases — tickets with ambiguous signals, security implications, or escalation needs that are not immediately obvious
• STEP 05Built the knowledge base — created reference articles for each ticket category the agent can match and surface in its response
• STEP 06Built the simulation interface — designed the triage workspace where users make their own classification decisions before seeing the AI output
• STEP 07Added the scoring system — tracks correct classifications, total triaged, and accuracy rate across the full ticket queue
• STEP 08Deployed to GitHub Pages and tested across all 22 tickets including all 4 edge cases

Evaluation Results

The agent was evaluated across all 22 tickets — 18 standard tickets and 4 edge cases. The simulation is designed so you can compare your own triage decisions to the agent's output in real time, building your own help desk judgment alongside the AI evaluation.

The edge cases were specifically designed to expose the limits of simple rule-based classification and demonstrate where AI reasoning adds value over keyword matching — particularly in tickets that combine multiple issue types, involve user frustration signals, or have security implications embedded in an otherwise routine request.