What an AI job agent is

An AI job agent is a system that pursues a job-search goal over multiple steps it chooses itself. You give it a target — the kind of role you want — and it decides what to look at, what to discard, and what to do next, without you approving each step.

That last clause is the whole distinction. A job alert runs a query you wrote and emails you the results. An agent decides what to search for, judges what it found, and acts on that judgement. One executes your instructions; the other makes decisions on your behalf.

The distinction matters commercially as well as technically. Something making decisions on your behalf deserves questions about what it can do without asking, and something running a saved query does not — so knowing which you have changes what you should check before trusting it.

Every serious implementation has the same four parts, and reading any product claim against them tells you quickly what is actually being offered.

Most tools marketed as job agents do the first two well and stop. That is genuinely useful, and it is a search-and-ranking product rather than an agent — which matters because the hard problems, and the risks, all live in the second half.

The fourth is the rarest by a wide margin. Learning requires outcome data, and outcomes in a job search are sparse, delayed and frequently never observed at all, which is why so few systems close the loop properly.

• Find — gather openings from sources, continuously rather than once
• Judge — decide which are worth your time, and explain why
• Act — tailor an application and submit it, or prepare it for you
• Learn — change what it does based on what happened last time

A job alert is a stored query. It cannot tell you that a role is a poor fit despite matching your keywords, because it has no model of what you are actually suited to — only the words you typed.

An agent reads the description and reasons about it. That is why an agent can reject a posting that matches every keyword and flag one that matches none of them, and why its judgements can be explained and argued with rather than simply tuned.

That ability to be argued with is underrated. A filter that excluded a role gives you nothing; an agent that says it skipped a posting because the seniority was two levels above your evidence gives you something you can disagree with, and disagreeing is how the thing gets better.

It cannot reliably represent you. An agent writing about your experience is working from what you gave it, and when that runs out it fills gaps plausibly rather than accurately. Every serious implementation therefore puts a human between generation and submission.

It also cannot get around the fact that applications are increasingly screened by other software. Volume was never the bottleneck it appeared to be, and an agent that makes volume cheap does not by itself make outcomes better.

And it cannot make the decisions that depend on your life. Whether to leave a stable job, whether a pay cut is worth a better role, whether to relocate — these need information the agent does not have and should not infer, and a system that decides them quietly is overstepping rather than helping.

The honest value is in the parts of a job search that are mechanical and that people do badly because they are tedious. Watching many sources continuously, noticing a role posted this morning, keeping track of what was sent where and what needs chasing.

Those are real gains and they compound over a search lasting months. A candidate who never misses a deadline, never loses track of which CV version an interviewer holds, and sees relevant roles on the day they appear is meaningfully better positioned than one juggling it manually.

It is a smaller claim than the marketing makes and a more defensible one. The agent is not getting you the job; it is making sure you are in a position to get it.

An agent is only as good as the profile it reasons from, and this is the part most people under-invest in. A parsed CV gives it your history; it does not give it your intent, your constraints or what you would refuse, and without those it optimises for continuity with your past.

That is why agents so often surface more of what you already did. Nothing in a CV says you are leaving that field, want less management responsibility, or will not relocate — so a system inferring a target from history will keep recommending the career you are trying to change.

The fix is cheap. State the target explicitly, state the hard constraints, and correct the first few judgements it gets wrong. Twenty minutes of that is worth more than any amount of tuning, because it replaces inference with fact.

• The roles you are actually targeting, not the ones you have held
• Hard constraints — location, right to work, notice, minimum salary
• What you would refuse, which never appears in a CV
• Corrections on its early judgements, so it learns your taste
• Which employers to exclude, and why

Autonomy sells. A saved search rebranded as an agent commands more attention and a higher price, so the label has spread far beyond systems that make decisions.

The test worth applying to any product: can it choose to do something you did not ask for, in service of the goal you gave it? If not, it is a very good search tool — which may be exactly what you want, and is worth buying under its real name.

Ask the follow-up too. If it can act, what specifically can it do without asking, and what happens when it gets one wrong? A product that answers both clearly has thought about it; one that treats the question as unusual has not.