Checking an agent's work before you rely on it
Here's the thing nobody puts in the marketing: agents are confidently wrong on a regular basis. They don't hedge, they don't flag uncertainty, and their mistakes read exactly like their correct answers. That doesn't make them useless — it makes verification part of the job. This guide is the routine.
Why agents make mistakes
A quick understanding helps you check smarter. Agents generate text based on patterns, not by looking things up in a fact database. When you ask a question, the agent produces the most plausible-sounding answer — which is usually right, but not always, and it has no internal alarm bell for the difference. The most dangerous errors are the plausible ones: a quote that's almost right, a date that's off by a year, a step in a process that doesn't quite apply to your situation. Keep that in mind and you'll check the right things.
The verification routine
Run through these whenever an agent's output matters — the bigger the stakes, the further down the list you go.
1. Read it fully, once
Skimming is how errors survive. Read the whole thing. You're checking for the obvious first: does it answer the question you asked? Did it follow your constraints? Is the tone right for the audience? A surprising number of problems are caught right here — the agent answered a slightly different question than the one you meant.
2. Check the checkable facts
Names, dates, numbers, prices, addresses, claims about what something does — anything that is either true or false. Verify these against an independent source: the actual website, the actual document, a quick search. Pay special attention to specifics the agent volunteered. If you asked for "a good plumber" and the agent named one with a phone number, that phone number needs checking. Agents invent plausible details more often than people expect.
3. Spot-check the reasoning
For advice, plans, or analysis, ask: does the logic hold up? Does step three actually follow from step two? Would this work for my situation, or is it generic advice wearing a specific hat? A useful trick: ask the agent to explain its reasoning for one key claim — "Why did you recommend this one over the others?" — and see if the explanation is solid or hand-wavy.
4. Test it against what you know
You're the expert on your own life and work. If something in the answer feels off, trust that feeling and dig in. "That doesn't sound right" is a valid reason to verify — agents don't have your context, and half their errors come from filling in context they don't have.
5. Ask for sources, then check the sources
When facts matter, ask the agent to cite where each key claim came from. Then open the sources. This is the step people skip, but it's the one that catches the serious errors — because sometimes the cited source doesn't say what the agent claims it says, or doesn't exist at all.
What to verify first
If you're short on time, check in this order of risk:
- Anything you'll act on. Money, health, legal, travel — if a decision rides on it, verify it.
- Anything you'll send to someone else. An error in an email draft is embarrassing; an error in a client proposal is expensive.
- Numbers and quotes. Prices, statistics, measurements — these are the most commonly mangled and the easiest to check.
- Specifics you didn't provide. Anything the agent added that you didn't ask for is unverified by definition.
When to redo it yourself
Sometimes checking costs more than doing. Redo the work yourself when: the task is short enough that doing it takes less time than verifying (a two-sentence email); the stakes are high and the topic is outside your knowledge (you can't check what you can't evaluate — get a human expert instead); or the agent has been wrong twice on the same task (a third attempt usually just costs you more time).
The goal isn't to distrust everything. It's to build a habit: agent drafts, you decide. Five minutes of checking turns a fast-but-unreliable assistant into a fast-and-reliable one — with you as the reliability.
Next: staying safe with agents — permissions, data sharing, and the guardrails worth setting up.