Insights: How I work
How I check an AI agent's work without redoing it.
Handing work to an agent only saves time if checking the result takes less time than doing the task. Here is how I keep the review short and still catch the mistakes that matter.
Why is review the hard part?
Getting an AI agent to produce something is easy. Knowing whether that something is right is the real work. If I have to read every line, rerun every lookup, and rebuild every list to trust it, I did not delegate anything. I just added a step.
So I treat review as part of the design of the task, not something I figure out after the result shows up.
What do I ask for back?
Before an agent starts, I decide what the finished report has to contain. Usually that is three things:
- The result itself. The draft, the list, the updated record, with a link to where it lives.
- How it knows. The source for each fact it is relying on. A file, a page, a record. Not "I found that."
- What it did not do. Anything it skipped, could not reach, or was unsure about.
The third item matters most. An agent that says "I could not open that page, so this part is missing" is far more useful than one that quietly fills the gap with something that sounds right.
What do I actually check?
I do not read everything with the same care. I sort the output into two piles.
The first pile is anything that leaves my hands: a message to a person, a post that goes public, a change to money or a contract. I read all of it, every time, before it goes anywhere. That rule does not relax as the agent gets better.
The second pile is internal work, like a sorted list, a cleaned up record, or a summary for me. There I spot check. I pick a few items at random and trace each one back to its source. If those hold up, I accept the batch. If one is wrong, I check more before I trust any of it.
What are the red flags?
A few patterns tell me to slow down:
- A number with no source next to it.
- A name, quote, or date that I cannot find in the material the agent was given.
- A report that says everything worked and lists nothing it skipped.
- A result that matches what I hoped for a little too neatly.
None of these prove the work is wrong. They tell me where to look first.
How do I make the next review shorter?
Every mistake I catch turns into a line in the task instructions. If an agent invented a source once, the instructions now say to name the source or say it is missing. If it claimed something was finished without checking, the instructions now say what "done" has to look like and how to verify it.
The goal is not a perfect agent. It is a task written so tightly that the common mistakes cannot happen, and the rare ones are easy to spot.
What is the actual lesson here?
Decide how you will check the work before the work starts. Ask for sources and for what was skipped. Read everything that goes out under your name, spot check the rest, and turn every catch into a better instruction. That is how delegation actually saves time.