AI, built and explained.
Deep dives on how modern AI works: the models, the agents built on them, and the engineering that gets them into production. What ships, what breaks, and why.
Evals Before Features: How to Test a Strands Agent You Cannot Unit-Test
Write the eval cases before the prompt, run them on every commit with a fake model and on every model change with a judge, and let a threshold gate decide the release.
Read the post →
How to Choose an Agent Control Flow Pattern: Loop, Router, Graph, Planner/Executor, Supervisor
Pick the pattern based on the task at hand, because each one decides who chooses the next step and what crosses each boundary.
Why Your Agent Loops Forever: Put the Review Step Outside the Executor
An agent loops when nothing in the loop can decide success or failure; a deterministic check on system state, reporting only to the planner, gives it an exit
A Prompt Is a Layered Program: Mapping Prompt Engineering Techniques to the Layer They Change
Every prompting technique edits one layer of the request, and every model or provider switch breaks the layers whose delivery differs
Understand it well enough to build it, or to know when not to.
- 01
Explained
How a model, technique or system actually works, with a diagram drawn for the post.
- 02
Built
When there is code, it runs: snippets you can paste and a public repo, MIT licensed.
- 03
Taken apart
Teardowns of real AI products and failures: what happened, why, and what to do instead.
New posts in your inbox. No digest, no roundup.
Free. Subscribing also sends you the AI in Production Playbook: checklists for the seven jobs every production agent needs done, from context to cost.
By subscribing you agree to Substack’s terms of use, privacy policy and information collection notice. Your address goes to Substack, which sends the emails. This site stores nothing. Privacy notice