LLM demos are easy. Shipping agentic systems that don’t break is not.

Once an agent is allowed to call tools, trigger workflows, or modify state, small probabilistic quirks turn into real engineering problems. In this talk, Luise walks through concrete failure modes she has encountered when integrating LLMs into real systems:

- structured output drifting just enough to crash downstream logic
- agents selecting the wrong tool or repeating calls
- reflection loops exploding token usage and latency
- prompt tweaks causing silent regressions
- context windows filling up with irrelevant history

For each failure mode, she will show code examples and the mitigation patterns that actually work: contract layers, validation and repair loops, tool scoping, iteration caps, cost guards, and lightweight evaluation harnesses.

This is a talk about what breaks when agents leave the playground and how to design systems that survive that transition.