In this session I will share **best practices** for using MCP setups and AI **Agents** in real development workflows 🤖. The goal is to improve both **quality** and **security**, since even with all the excitement around these tools, it’s still very easy to create confusing interactions, accidental data exposure, or Agents that trigger actions we never intended. I will also present **reports comparing how different LLM models generate code**, showing clear differences in stability, output consistency, and the amount of guidance each model needs to avoid mistakes. Along with that, I’ll bring recent **figures showing the surprisingly low productivity gains** many teams report when adopting AI tools—often just small improvements instead of the big jumps everyone expects. We will look at why this happens: unclear prompts, weak tool design, loose permissions, and wrong assumptions about what the models can really handle. The talk focuses on practical habits that developers can apply right away: tightening tool access, simplifying and focusing context, adding lightweight validation steps, and using safety patterns that prevent the common failures. My goal is to help teams build **more reliable, productive, and secure AI workflows**, making AI code generation a real benefit rather than a risky experiment 😅.

Talk Level:
INTERMEDIATE

Bio:
International Speaker, JavaChampion, Cofounder of JBCNConf and DevBcn conferences in Barcelona, and AI4Devs conference in Amsterdam. Currently working as a Staff Developer Advocate in Java at Sonar (SonarQube), focused on Code Quality, Dev Productivity, AI & Security. I have worked as a (paid) developer for more than 30 years ago using multiple languajes, but for the last 15 using Java. Although I started when I was 14 with my Amstrad CPC 6128 :) I am very interested in simulated reality, psychology, philosophy, and Java.