architecture

Not better function calling — decoupling. When tools expose MCP servers, any agent can compose any system freely. The heterogeneity problem becomes a configuration problem.

We called them language models and spent years confused about why they could reason. The name stuck to the interface, not the mechanism.

When behaviour emerges from both code and model responses, reading source files isn't enough. You debug by examining execution traces.

Tool enthusiasm is often net-negative. Context pollution degrades performance faster than features improve it. The principles that actually work.

The instinct to add a rule to a skill file is usually the wrong abstraction. A CLI wrapper enforces at the tool level: zero deliberation, zero token cost.

LLM instruction files are code. They have the same failure modes — with one interesting twist that changes everything.

AI agents know how to start a task. They don't always know when to stop. The contract pattern is the architectural fix.

Building AI systems that work in the real world requires thinking beyond the demo. What actually matters when users depend on your models.