No. 11 · Agents · Latest
Grade the outcome, not the transcript
An agent can say all the right things and still leave the database wrong. Check the state it leaves behind, and you get a test you can trust and a reward you can train on.
Read the essayTarkika · EssaysFree, every Sunday
The SundayLetter
One essay a week on building AI that decides: what worked, what failed, and the code. No noise, and you can leave any time.
No. 11
Grade the outcome, not the transcript
Agents26 Oct 2026 · 5 min No. 10Read-only first: how much access should an agent get?
Agents24 Oct 2026 · 5 min No. 09Your review queue is a budget
Decisions19 Oct 2026 · 5 min No. 08Log every decision your agent makes
Agents18 Oct 2026 · 5 min No. 07Two decision models in one day
Decisions11 Oct 2026 · 5 min No. 06The Jevons paradox of decisions
Decisions4 Oct 2026 · 6 min No. 05Act, review, or escalate: agents that know their limits
Agents27 Sep 2026 · 6 min No. 04Six ways to make one decision
Decisions20 Sep 2026 · 6 min No. 03Your LLM's 0.95 isn't a probability
LLMs13 Sep 2026 · 6 min No. 02When 0.8 should really mean 80%
Calibration6 Sep 2026 · 8 min No. 01Why 0.5 is almost always the wrong threshold
Decisions30 Aug 2026 · 8 minNext
What I wish I knew before my first AI internship
CareersComing soonStart here
New to the letter? Three steps in.
1
Read the night shift
The book's prologue, free: one analyst, two hundred alerts, and the night that goes differently.
Read the prologue 2Then this essay
Why one line at 0.5 lets threats through, and how costs draw better lines.
Read the essay 3Get it every Sunday
One essay a week, with the code, in your inbox. Free.
Subscribe
From the book
Decide, Don't Generate
The ideas behind these essays, in 22 chapters with runnable Python labs. On Kindle and in paperback. Most of what an AI agent does isn't writing. It's deciding.
The Sunday Letter