The 30-Minute Weekly Review That Keeps an AI Agent Honest
An autonomous agent fails in two directions, and they need two different remedies.
It can stall — blocked, confused, waiting for you. That failure is loud, and ten minutes a day catches it. The other failure is drift: the agent works beautifully, every day, on things that don't matter. Drift is silent. It never shows up in a daily check-in, because every individual day looks productive. It only becomes visible when you zoom out — which is exactly what the weekly review is for.
The daily ritual asks: "is it moving?"
The weekly review asks: "is it moving somewhere I want to go?"
Before you start: read, don't chat
The review begins in the files, not in a conversation. Ask the agent to summarize its week and you'll get a summary shaped to please you — agents are agreeable, and a chat transcript is a sales pitch. The files are the record. Read them cold:
- The work journal (the last 5-7 daily entries) — what actually happened, including what broke.
- The metrics file — the one number. Up, flat, or down.
- The decisions file — what did the agent decide on its own this week? This is the highest-value read in the whole review, and the one people skip.
- The backlog — what's at the top now, and what's been sitting there for three weeks pretending to be a priority.
Fifteen minutes, tops. Then, and only then, open a session.
The four questions
1. What did the agent decide that I would have decided differently?
Not "was it wrong" — different. Every gap between its judgment and yours is a bug in the job contract, and it's a bug that will repeat every week until you fix it in writing. This is the mechanism by which an agent gets smarter without you ever saying the same thing twice.
The reflex to build: fix behavior in the contract, not in the chat. Correcting the agent in conversation feels efficient and buys you exactly one session of compliance — the conversation dies, and the correction dies with it. One line added to the contract buys you every session from now on. Same effort, permanent effect.
2. Is the top of the backlog still the top?
Priorities rot. Last week's obvious next thing is often this week's obvious waste — because the numbers moved, because a bet failed, because you learned something. The agent will not spontaneously re-rank its backlog against your strategy; it takes the top item and starts working. If the top item is stale, you get a week of excellent work in the wrong direction, delivered on time.
Re-ranking takes two minutes. It is also the entire steering wheel of the operation. Treat the top three items as your budget: whatever sits there is what the week gets spent on, so put the expensive attention there deliberately.
3. Did the one number move — and does the answer change anything?
Read the one number, and answer honestly:
- Moved up: what caused it? Ask the agent to find out, then do more of that specific thing. "Keep going" is not an answer; it's a way of not looking.
- Flat, and the fair trial still has time on it: do nothing. This is the hardest discipline in the system. Early flatness is not evidence — it's the ramp. Changing course here destroys the experiment and teaches you nothing.
- Flat, and the fair trial has run out: kill it, write the cause of death, move the budget. The date was set for a reason; honor it.
4. What is the agent quietly stuck on?
Look for the blocked items — the ones waiting on a decision, an account, a key, five dollars. A good agent parks these and works around them, which is exactly what you want operationally, and exactly what makes them easy to miss. Each blocked item is often sixty seconds of your time holding back days of the agent's. The weekly review is where you spend those sixty seconds — reliably the highest-leverage half-minute of the week.
The output: three edits, not a conversation
A weekly review that ends in a satisfying chat has failed. It has to end in file edits, or none of it survives to Monday:
1. Contract edit -> the correction that stops the drift repeating
2. Backlog re-rank -> the top 3 that define next week
3. Unblock -> answers written into the owner's inbox file
(+ a metrics conclusion, if the number said something)Then commit. That's the review — a half hour that decides where a week of tireless, cheap, non-stop work gets pointed.
The temptation to review daily
You'll want to do this every day. Don't. Steering daily produces a system that gets a new direction every 24 hours and finishes nothing — you become the drift. Agents work in weeks; the numbers that matter move in weeks. Ten minutes a day to unblock, thirty minutes a week to steer. The gap between those two is where compounding actually happens — and it's a gap you have to protect, mostly from yourself.
Do this now
Put a 30-minute block in your calendar, the same slot every week. In it, open exactly four files — journal, metrics, decisions, backlog — and write down one contract edit before you close them. If you find nothing worth editing in the contract, you didn't read the decisions file closely enough. There is always something.