We’re currently “programming” LLM systems in English: prompts, instructions, and policies written in mostly "English" but its dangerously ambiguous. If you want agents that scale, you need structured languages: schemas, typed tool contracts, state machines, and intent representations that can be tested, diffed, and optimized.

But is this a silver bullet? - This talk argues that natural language is great for communication, but poor for control. We’ll explore a practical middle ground exploring the uses of mechanistic interpretability to understand how a machine thinks and how we apply these learnings into building more robust systems.

What they’ll leave with:
- Pro's and Con's to language based LLM frameworks vs more stricter (like DSPy)
- Introduction to mechanistic interpretability and how you can better diagnose your LLM
- How this unlocks stable evals + self-optimization