Description of the Training

Generating code from a prompt works for prototypes. In safety projects it fails: no specification means no traceability, and no traceability means no evidence. Code without evidence is worthless in an audit.

This training shows the disciplined alternative. The specification comes first. AI agents derive architecture, code and tests from it, and an engineer reviews and approves after every phase. All artifacts are versioned as text in a Git repository, so every change the AI makes is traceable. The V-model stays the backbone; the agents take over the routine work.

An agent that reads your repository sends source code to a model as context. Which parts, which model processes them and where it runs are decisions you make in the architecture of your toolchain. The training shows how to limit the context to what is needed, exclude confidential modules or route them to a locally hosted model, and check AI output for license risks.

We demonstrate the workflow live on a real example: from a formal requirement to an automated validation test on a microcontroller on a HIL test bench. In the exercises you apply the workflow yourself on a sample project.

Objectives and Benefits

Write specifications that an AI agent can implement unambiguously and verify against.
Build the context for AI agents and limit it to what is needed: excerpts from standards, templates, project artifacts, skills.
Manage requirements, architecture and tests as code and check traceability automatically.
Exclude confidential parts of a project from the AI or route them to locally hosted models.
Check AI-generated code for license risks and document its origin in the repository.
Set review gates so that AI output arrives in reviewable units, and check it for typical errors.
Transfer the approach to a project of your own.

Agenda

Day 1 – Method and protection

From vibe coding to agentic engineering: the limits of prompt-based code in safety work
Spec-driven development: specify, plan, tasks, implement, with approval gates
Context engineering instead of prompt engineering: what the AI needs to know for reproducible results, and what it should not see
Everything as code: requirements in TRLC, traceability with LOBSTER, architecture and design as text, consistency checks in CI/CD
IP protection in the toolchain: what an agent reads and where it sends it, exclusion rules per repository, separate repositories for confidential modules, choosing the model by protection need, up to a locally hosted model
Your context is your know-how: version skills, templates and prompt libraries and protect them with access rights, like source code

Day 2 – Practice

AI agents along the V-model: from SWE.1 requirements to SWE.6 validation test
Live demonstration on a HIL test bench: a new feature from requirement to passed test
Loop engineering: the agent iterates against predefined tests until all acceptance criteria are met
Securing the results: license checks for AI-generated code, marking AI-generated changes in the Git history, keeping credentials out of the context
Exercise: specify and implement a feature in the sample project, with review gates, exclusion rules and model choice by protection need
The engineer's role: from writing to reviewing and designing; building technical depth in the team on purpose

Target Audience

Software developers, software architects, test engineers, team leads and technical project managers in safety projects.

Prerequisites

Experience in embedded software development (C), basic knowledge of Git and functional safety. Please bring your own laptop for the exercises.

← All trainings