The travel assistant of Nordlicht Reisen GmbH (R-01) talks to customers, reads third-party documents and hotel descriptions, generates destination images (R-05) and can call tools. Refunds, legal questions and blocked accounts always go to a person (R-02). This lesson sets out how that is evidenced in operation: with guardrails, a safety evaluation before every version, traceable traces and clear oversight modes per tool. Paper plus two compiled sketches (Bicep, C#); all data is made up (R-12). Cost drivers: moderation and evaluation runs by consumption, telemetry per gigabyte.
Prior knowledge. Treating foreign text as data is covered by "AI track, lesson T6.2", content filters and Prompt Shields by "T6.3", hard rules in code by "T6.4", approvals by "T7.3", rights per tool by "T8.3", evaluators and the judge by "T9.2 and T9.4", evaluation in the build by "T9.5", traces by "T10.2". None of that is repeated here. New is the Foundry side: guardrails on agents, the hosted evaluators, provenance of images and the question of who supervises the agent.
The requirement
All numbers are assumptions of the task.
| No. | Requirement | Type | Measure | Weight | Source |
|---|---|---|---|---|---|
| 1.5-01 | Inputs and answers are checked for harm and attacks. | F | all requests | Must | Security |
| 1.5-02 | For the agent, tool calls and tool responses are also checked. | F | 2 additional intervention points | Should | Security |
| 1.5-03 | A safety evaluation runs before every new version. | R | 40 cases, 0 failed | Must | Quality |
| 1.5-04 | Every action is traceable: version, source, approval. | R | 100% of actions | Must | Audit |
| 1.5-05 | Generated destination images carry provenance marks (R-05). | F | all images | Should | Editorial |
| 1.5-06 | Actions with external effect need an approval. | R | 0 without approval | Must | Legal |