Notes, mostly to myself.
What I've figured out building with agents. What worked in production. What didn't. Published in case it's useful to someone else.
Why we replaced RAG with an LLM-native wiki
For a year our agents retrieved. Then we let them reason over a curated knowledge base instead — and most of our reliability problems quietly disappeared.
Running a company as a fleet of stateful agents
Domain-based AI roles instead of Scrum. How Temporal coordinates long-lived agents across billing, operations and support — and what that does to an org chart.
Opus, Sonnet or Haiku: a cost & latency model
A practical framework for model selection under real rate limits, latency budgets and unit economics — from someone who pays the bill.