Every team has now watched an AI agent scaffold a widget beautifully, then three prompts later invent a package that doesn't exist, rewrite state management it was never asked to touch, or quietly break a passing test. The tooling isn't the problem. What's missing is context engineering — the discipline of deciding what the model sees, what it remembers, when it stops, and who checks it.

This session is about that discipline, drawn from building 40 hours of AI curriculum for a university programme and training enterprise teams to use these tools on real work rather than in demos.

You'll leave with:

Why context — not prompt wording — becomes the dominant variable the moment you move past a toy example
How to scope a task so an agent works on a bounded slice of a codebase rather than the whole repo
How to design verification loops: what gets checked automatically, what needs a human, and where to put the stop condition
The four failure modes that show up again and again, and the specific guardrail for each
How to write a prompt a teammate can reuse next month without you standing next to them

This is not a tour of AI tools. It's the operating discipline that decides whether they save your team time or quietly cost it.