Nordlicht Reisen GmbH (AI-103, R-01) has received the folder from lesson 5.1. Before anyone releases money for the team there, a second person reads it critically and asks: Why these agents, what happens if it is not right? This lesson practices defending, shows openly what the course has not taught, and turns the "open" places into a checklist. It is paper: No command runs against Azure; at the end stands the acceptance that two examiners read.
Prior knowledge. Lesson 5.1 (folder, sheet A-01 to A-08, conflicts K-1 to K-6, sketch) and the decision notes of lessons 1.1 to 4.7. The form of the note is taught in "K4, lesson 4.1.3". The map of the exam skills is in lesson 0.1; here it is read against what the course has delivered. "AI-103, lesson 6.2" has the same form.
New in this lesson. No new facts of the case. New are the questions of the second person, the gap list and the checklist; they summarize what the lessons say and change nothing about M-nn, F-nn or the resolutions from 5.1.
The answer in three parts
A question is not answered with a preference. The pattern has three parts: the requirement (which row of the sheet), the decision (what was chosen, with the lesson) and the limit that flips it (at which measured value or which fact the answer would be different). If the limit is missing, it is an opinion. If the requirement is missing, something was built that nobody ordered.
| Area | Question | Answer in the pattern |
|---|---|---|
| 1 | Why seven agents and not one? | A-01, A-02 → permissions and fields differ, only two roles have A2 (1.1) → flips for clearly bounded cases without strict separation: Then one agent with tools suffices (so Learn) |
| 1 | Why build our own instead of Foundry's memory service? | A-02, A-07 → index gedaechtnis with filter mandantId (1.3, 2.4) → flips if Learn lists the service as stable and the separation per call is proven |
| 2 | Why no fine-tuning? | A-04, A-06 → four conditions, until then prompts and examples (2.2) → flips as soon as they apply; a fine-tuned model is another deployment and changes R-08 |
| 3 | How do you know that a new version is not worse? | A-04 → gate with four numbers (4.7), 20 of 60 cases by two reviewers (3.1) → flips if the spread of the set overtakes the tolerance; then the gate changes in code with review |
| 4 | Why are guardrails not enough? | A-01 → the service catches harm and attacks, the code the business limits FG-1 to FG-4 (4.4, 4.5) → never flips for hard rules: A guardrail does not decide whether an agent may call a tool (F-05) |
| 4 | Why the canary in code? | A-04 → Learn is contradictory on percentages for prompt agents (open, 4.6) → flips as soon as one side documents them for prompt agents |