Building an IT Support Agent — RAG + Parallel Tool Calls

My second AI agent goes beyond the calendar scheduler.

It diagnoses IT issues, searches company documentation, checks system status, and fixes problems — all in one agentic loop. Two new concepts: RAG and parallel tool calls.

---

What It Does

Employee types: “My VPN is not working and I can’t access Jira”

Turn 1: LLM → search_knowledge_base("VPN not working")
              search_knowledge_base("Jira access")     ← parallel!

Turn 2: RAG returns policy docs → LLM →
              check_vpn_status(EMP-4521)
              check_tool_access(EMP-4521, "Jira")      ← parallel!

Turn 3: Both expired → LLM →
              enable_vpn(EMP-4521)
              enable_tool_access(EMP-4521, "Jira")     ← parallel!

Turn 4: Both fixed → LLM →
  "Fixed both! VPN renewed until Aug 22.
   Jira access restored. ✅"

---

New Concept 1 — RAG First

The system prompt forces the agent to search docs before taking any action:

ALWAYS call search_knowledge_base first.
Never check status without reading policy.
Never enable access without checking status first.

This grounds every decision in real company policy. The agent doesn’t guess — it reads, then acts.

---

New Concept 2 — Parallel Tool Calls

In the calendar agent, tools ran one at a time. Here, the LLM calls multiple tools in the same turn:

// LLM response — two tools in one turn
content: [
  {
    type: "tool_use",
    name: "check_vpn_status",
    input: { employee_id: "EMP-4521" }
  },
  {
    type: "tool_use",
    name: "check_tool_access",
    input: { employee_id: "EMP-4521", tool_name: "Jira" }
  }
]

// Your code handles both
for (const toolCall of toolCalls) {
  const result = await executeTool(toolCall.name, toolCall.input);
  toolResults.push({ type: "tool_result", ... });
}

Both tools execute, both results feed back to LLM in one shot. Fewer turns = faster resolution.

---

The 5 Tools

1. search_knowledge_base  ← RAG — always first
2. check_vpn_status       ← read current state
3. enable_vpn             ← fix if expired
4. check_tool_access      ← read current state
5. enable_tool_access     ← fix if expired

Pattern: search → check → fix Never fix without checking. Never check without reading policy.

---

What’s Different from Calendar Agent

	Calendar Agent	IT Support Agent
RAG	❌	✅
Parallel tools	❌	✅
Tools	2	5
DB	None	None (JSON RAG)
Input	Date + time	Employee ID + issue

Same agentic loop underneath. Two new concepts layered on top.

---

Key Lesson

The system prompt is the agent’s personality.

"ALWAYS call search_knowledge_base first"

This one line changes the entire behavior. The LLM follows it every single time. System prompt design is as important as code.

Next: why vocabulary mismatch is the real reason to add vector embeddings to RAG.