LLM agents in production: what actually works

Demo post. Replace this text with your own material — files live in src/content/blog/, one post = one markdown file.

Why agents break

Most agent pipelines die not at the model, but at the boundaries: between tools, between steps, between user expectations and reality.

What to do

1. Less autonomy, more checkpoints.
2. Log every decision the agent makes.
3. Evaluate on live traffic, not synthetic data.