The lesson that protects you
Everything so far has been about what agents can do. This one is about what happens when they cannot, which they will not tell you, because they do not know.
None of what follows is a reason to avoid agents. It is the difference between people who use them well and people who tell a slightly haunted story at a conference eighteen months later.
One: the confident wrong turn
An agent hits an ambiguity you never thought about. The file has two tabs. Two clients share a surname. Last quarter could mean the calendar one or the financial one.
A person stops and asks. An agent picks. It picks reasonably, it picks quickly, and it does not flag that a choice was made at all. The rest of the work is then built perfectly on top of the wrong tab.
YOU SAID WHAT IT MIGHT HAVE PICKED
last quarter calendar Q, not your financial year
the client file the older duplicate in the archive
everyone on the team including the two who left in March
this month's numbers the draft version, not the signed off one
the usual format whatever format it saw most recently
urgent tickets its own idea of urgent, not your rulebook
The tell is that the output looks immaculate. Wrong work does not arrive scruffy. It arrives beautifully formatted, on time, and built on a foundation you never inspected.
Two: silent failure
This one is worse, because there is nothing to see. The agent could not read a file, so it worked with what it could read. A search returned nothing, so it went on without it. A record was locked, so it skipped it.
Sometimes it will mention this. Often it mentions it in one line, in the middle, and the summary at the end says the task is complete. Which, from where it is standing, it is.
The most useful sentence you can add to any agent instruction: tell me explicitly what you could not access, could not find, or skipped, and put it at the top rather than the end. You will be surprised how often that list is not empty.
Three: small errors that compound
In a one step task a small error is a small error. In an eight step task, step one feeds step two. A slightly wrong list at the start becomes a confidently wrong report at the end, and every step in between looked fine.
Step 1 Pull the list of active customers
(it included 6 who cancelled last month)
Step 2 Calculate average revenue per customer
(denominator now wrong, number looks plausible)
Step 3 Compare against last quarter
(comparing a wrong number to a right one)
Step 4 Write the summary
(revenue per customer is down, we should investigate)
Step 5 Draft the board slide
(a fictional decline, beautifully typeset)
Nobody made a big mistake. The first one was six rows.
This is why the checking advice from the last lesson matters so much. Check the first step, not the last one. The last one always looks great.
In plain English
- Confabulation:
- Producing something plausible and specific that is simply not true. Confidence is not evidence.
- Silent failure:
- The agent could not do part of the job, carried on regardless, and reported success.
- Compounding error:
- A small mistake early that every later step builds on, growing as it goes.
- Irreversible action:
- Anything you cannot undo: sent, published, paid, deleted, agreed.
- Blast radius:
- How much damage the worst version of this task could do before anyone noticed.
Four: the irreversible thing
Every failure above is survivable. This one is the category that keeps people awake.
An agent has no felt sense of consequence. Sending an email and saving a draft are the same size of action to it. Deleting forty files and deleting one are the same shape of job. It cannot feel the difference between a message to your colleague and a message to your biggest customer, because there is no difference in the mechanics. The difference is entirely in your world, and it does not live in your world.
SEND Anything that lands in someone else's inbox or phone
PUBLISH Anything a stranger could see and screenshot
PAY Anything that moves money in any direction
DELETE Anything without an obvious undo button
AGREE Anything that commits you to a date, price or promise
These are not beginner restrictions you graduate out of.
Experienced teams keep every one of them.
Checkpoint
The four failure modes: confident wrong turns, silent failure, compounding errors, and irreversible actions. The first three cost you rework. The fourth costs you a phone call you would rather not make.
Find, then decide
Most risky agent tasks have a safe version hiding inside them. The trick is to split the finding from the deciding.
โ Weak prompt
Prompt
Go through the customer list and clean it up. Remove anything that looks like a duplicate or out of date.
Output
Cleaned 1,204 records. Removed 47 duplicates and 112 records with no activity in 12 months. List is now tidy.
Looks like a duplicate is a judgement call, and remove is permanent. Two of those 112 dormant accounts were seasonal customers who order every November. Nobody will notice until November.
โ Good prompt
Prompt
Go through the customer list. Do not delete or change anything. Produce a separate document listing records you believe are duplicates, with both record IDs and why you think so, and records with no activity in 12 months. I will decide what happens to them.
Output
No records changed. Report produced: 47 possible duplicate pairs with reasons, 112 dormant records with last activity dates.
Same work, same time saved, none of the risk. The agent does the finding, which is the boring part, and you do the deciding, which is the part that has consequences.
What good practice actually looks like
You do not need to become suspicious of everything. You need three habits, and they take almost no time.
Check the input, not just the output. Spot check the first thing the agent gathered. If the starting list is right, the rest usually follows. If it is wrong, nothing after it matters.
Ask for the exceptions. Make the agent report what it skipped, could not find, or had to guess. A run with no exceptions on a messy task is a warning sign, not a triumph.
Keep the five irreversible actions yours. Draft, do not send. Suggest, do not publish. Flag, do not delete. This costs you a few seconds and removes almost every genuinely bad outcome.
There is a useful mental model here: treat an agent as a brilliant, tireless intern on their first day, who has read everything and understood the office politics of nothing, will never admit to being confused, and has somehow been given the keys to the filing room. You would not let that person send client emails unsupervised. Same policy, same reasons.
๐ Quiz
Question 1 of 4What is a confident wrong turn?