“Agent” has become one of those words that means whatever the person using it needs it to mean. It is worth stripping back, because underneath it is a real and fairly simple distinction.
An assistant answers when asked. An agent takes steps without being asked for each one.
That is the whole difference. Everything else — the demonstrations, the diagrams, the vocabulary — is detail hanging off it.
The difference in practice
Ask an assistant to look at the electricity bills and it will look at the electricity bills and tell you what it found.
Ask an agent the same thing and it may notice the standing charge changed in March, go and find the notification letter you missed, check whether the change matches what the letter said, and tell you that it does not. Nobody asked for steps two, three and four. It took them because they followed from the first.
That is genuinely useful and it is also the reason agents need to be constrained carefully, because the same willingness to take an unasked-for step is what makes a mistake propagate.
An honest day
What this looks like in an ordinary week, rather than in a launch video.
Early. It reads what arrived overnight. Nothing dramatic — sorting, noticing, and drafting a short written pass on what needs a person.
Mid-morning. You ask a question with an answer buried across several documents. It reads them, gives you the answer, and shows you which documents it came from.
Around lunch. It notices something you did not ask about: an invoice that duplicates one from August, a renewal in three weeks that costs more than last year. This is the part people find unfamiliar, because nothing prompted it.
Afternoon. You ask it to draft three replies. It drafts three replies. They are eighty per cent right and you change the parts that need judgement.
Evening. Nothing. Which is correct; an agent with nothing useful to do should do nothing, and one that manufactures activity to look busy is worse than one that sits still.
Nine-tenths of the value is in the noticing, and almost none of it is impressive to watch.
What it does badly
Being clear about this matters more than the capability list, because the failures are predictable.
Anything requiring a judgement it cannot check. It can tell you an invoice looks duplicated. Whether to challenge the supplier depends on the relationship, and it does not have the relationship.
Long chains where an early step was wrong. The characteristic agent failure. Step one misreads a date; steps two through six are all diligent, all correct in their own terms, and all wrong. The more steps it takes unprompted, the more this matters, which is why a good agent shows its working.
Knowing what it does not know. Given a question its material cannot answer, the honest response is “this is not in what I have”. Systems tend to produce something plausible instead. Any agent worth trusting has to be built to say no, and you should test that deliberately.
The thing that makes it work at all
None of the above is possible without the material. An agent with no access to your documents can only take steps in the world of general knowledge, which for your own affairs means it can take no useful steps at all.
Which is why the interesting question about agents is rarely the agent. It is what it has read — and whether reading it required handing your archive to somebody else.
How to judge one
Three questions, in order of usefulness.
Can it show its working? Not a summary. The actual steps, and the documents each one used.
What can it do without asking? Reading and drafting, ideally. Sending, paying and deleting should require you every time, no matter how confident it is.
What happens when it is wrong? If the answer is that you find out later from the consequences, the design is wrong regardless of how good the demonstration was.
What it should and should not be allowed to do →
What this looks like in a household →
Local AI. Private data. Local training.
$8,995 the first year — everything included. $4,995 each year after. The first year costs more because it contains the Jetson Orin placed and configured, your archive loaded and the first training run; every year after is the service running.
Or add the Archive Assessment first — $1,950, credited in full
You order and pay at checkout; we write back within days — a decline returns every dollar, and until our letter confirms the year you may withdraw in writing. From that letter the year is final, and it arrives on the date the letter names. Prefer to write to us first?