Nordlicht Reisen GmbH is building the group-travel team (M-01 to M-03): seven agents, three MCP servers, a controller as a Container App (M-10). Nothing is chosen yet. Which agent gets which deployment? How many runs can the controller carry at the same time without blowing the quota? How will you tell later which customer triggered a run? And what does a developer work with on the first day? The lesson answers on paper and with a compiled C# example. You do not need an Azure account.
Prior knowledge. Model families and model card: "AI track, lesson T1.3". One model per task, Data Zone Standard, quotas, measurement plan: "AI-103, lessons 1.1 and 1.3". KEDA, the role of the rule and minReplicas: "K3, lesson 3.5.7". Traces and Application Insights: "AI track, lesson T10.2", "K3, lesson 3.4.2". New is the view of a team: every role needs a choice, calls multiply, and a trace has to hold together across agents.
New in this lesson. (1) The assignment of the roles to chat, chat-klein and embed; R-08 stays. (2) The controller limits itself: 3 runs at the same time per replica, 25 model calls per run, 5 replicas (assumptions). (3) Every span of our own carries mandantId, laufId, agent (M-11). (4) The repository gruppenreisen has a dev container and a file with AI instructions.
The requirement
All numbers are assumptions of the exercise.
| No. | Requirement | Type | Measure | Weight | Source |
|---|---|---|---|---|---|
| 1.4-01 | Every role runs on the smallest deployment whose results pass in the test set (M-12). | Q | 90% passed per role | Should | Cost (M-11) |
| 1.4-02 | A draft arises in under 10 minutes (M-02), even with several runs at the same time. | Q | 95% of the runs under 10 minutes | Must | Business unit |
| 1.4-03 | A run makes at most 25 model calls. | R | Counter in the code aborts | Must | Cost |
| 1.4-04 | Peak load produces no rate-limit errors. | Q | 0 errors in the test | Should | Operations |
| 1.4-05 | Every span of a run names customer, run and agent (M-07, M-11). | F | 100% of the spans | Must | Operations, chargeback |
| 1.4-06 | A new developer builds and tests without an Azure account. | Q | under 30 minutes to the first green test | Should | Development |
Cost frame. Cost drivers: tokens per model call (reasoning tokens count too), Container App compute time per replica, Application Insights per gigabyte of traces. Prices: the pricing pages of Foundry, Container Apps and Azure Monitor.