Capstone: an IT helpdesk agent

This is the whole course in one program. Deliberately boring, deliberately recognisable: the internal helpdesk that every company has.

Someone types "my VPN keeps dropping". Your agent looks up the policy, answers if it can, and raises a ticket if it cannot.

What it has to do

Every piece here is something you already built.

• Search. Find the right policy, the Module 10 idea, as search_knowledge
• Act. Raise a real ticket, as create_ticket
• Identity. Managed identity, no keys, from Module 7
• Trace and score. Watch it and grade it, from Module 12
• Ship. Pipeline, budget, guardrails, from Module 18

The settings from Modules 6 to 18 are already correct in the starter. Your job is the agent itself.

What to write

1. TOOLS, describing search_knowledge and create_ticket
2. The agent loop, with tools=TOOLS and deployment=AZURE_DEPLOYMENT
3. Feed every tool result back before the next chat

Point three is the one that catches people, and it is the same mistake from Module 2 and Module 12. Both tools can run perfectly and the answer can still ignore them.

How to know it is good

1. Does it loop?
2. Does it call both tools?
3. Does it use what they returned?
4. Is it traced and scored?
5. Does it fail safely when you break something on purpose?
6. Does the cost work out to a number you would defend?

The point

Every earlier module was one part. This is the assembled machine, and you have written every piece of it.

Break it on purpose

Drop the tool result messages. Both tools still run, the trace looks healthy, and the check fails. If you remember one thing from this course, make it that one.