A 20-minute weekly review that keeps automation useful

Bring one week of evidence

Before the clock starts, pull one week of the agent’s work into a single view: what completed, what failed, what is paused, and—most important—the outputs people accepted alongside the ones they corrected by hand. The corrections are the signal. They show you exactly where the agent’s output and your judgment still disagree.

Reviewing from memory is the common failure here. Memory keeps the one dramatic miss and quietly drops the twelve small edits someone made every morning, so you end up fixing the wrong thing. A weak week even looks fine if you only remember its single good output. The trade-off is a few minutes of gathering, and it is worth paying: a review built on evidence points you at the real problem, not the memorable one. You know the view is complete when you can name the corrections, not just the wins.

The 20-minute agenda, in one view

The whole point is a review short enough that you keep doing it. Time-box twenty minutes into four blocks and stop when each one ends:

• Minutes 1–5 — Value: which outputs actually changed a decision or finished a task?
• Minutes 6–10 — Friction: where did people edit, wait for, or step in front of the agent?
• Minutes 11–15 — Risk: which permissions or connections are broader than the work needs?
• Minutes 16–20 — Act: make one improvement, remove one low-value thing, and write one sentence about why.

The failure mode is the open-ended review: it sprawls toward an hour, gets postponed to “when things are calmer,” and never happens again. A firm timer prevents that. Twenty minutes is admittedly shallow—you will not diagnose everything in one sitting—but a shallow review you run every week beats a thorough one you abandon after two.

Minutes 1–15: value, friction, and risk

The first fifteen minutes are diagnosis, not action—resist fixing anything yet. In the value block, separate “opened” from “used.” An output someone opened and then ignored is not value; only the ones that changed a decision or completed a task count. This is the most common self-deception in the review: an unread summary that technically “ran successfully” still failed at its only job.

The friction block reads the human corrections you gathered. Repeated edits in the same place mean the instructions or the input are wrong, not the person. The risk block is a quick access check: does any workflow still hold a connection or permission it no longer uses? Broad access that nothing needs is the quiet liability—harmless until the day it is not. Judge each block by whether you can point to a specific example, not a general feeling.

Minutes 16–20: change one thing, remove one thing

Now act, but narrowly. Pick the single smallest change with the highest expected value: narrow an input list, tighten an output template, mute a notification nobody reads, or add one approval where a mistake would leave the dashboard. One change per review is a discipline, not a limit—if you alter three things at once and results shift, you cannot tell which change did it, and next week’s review becomes guesswork.

Then spend part of the block removing work, not just improving it. Pause a schedule whose output went unopened, delete a memory that is no longer true, or drop a connection no active workflow uses. The trade-off worth naming: keeping a low-value workflow “just in case” feels safe, but every one you keep is attention you spend and access you maintain. Good operations includes subtraction, and the weekly review is where you do it.

Record the decision in one sentence

Close every review with one written line: “We changed X because Y; we expect Z by next review.” That sentence does two jobs. It settles the same debate before it can restart next Friday, and it turns your change into a testable prediction—next week you either saw Z or you did not, and either way you have learned something.

The failure is the vague note: “improved the report” gives future-you nothing to check against. Name the change, the reason, and the expected effect. It costs thirty seconds, and skipping it is why teams quietly rerun the same argument month after month. The record is what turns twenty separate reviews into a system that actually learns.