You run a search once, on a Tuesday, for a senior DevOps engineer. By Friday, without touching the tool again, you have a shortlist that didn't exist when you left the office. That's the pitch behind every "AI recruiting agent" landing page this year, and it's also the point where most recruiters stop trusting the term. An agent is not a smarter search bar. It's software that keeps working on a goal after you close the tab, and understanding exactly what it does in that gap is what separates a real time-saver from a black box you're afraid to rely on.
This is a plain breakdown of what an AI recruiting agent actually does between the moment you define a role and the moment a candidate shows up in your pipeline, what still needs a recruiter's decision, and where the DIY version (a custom agent you build yourself in ChatGPT or Claude) runs into a wall that a purpose-built one doesn't.
The short version: an agent takes your search criteria, keeps running against fresh candidate data on its own, and surfaces people worth a look. It does not decide who gets hired, and it should not be marketed as if it does.
What "AI recruiting agent" actually means
A search tool answers one query and stops. You type in a role, get a list of candidates, and if you want updated results tomorrow, you run the search again yourself. An agent is the same underlying search, minus the part where you have to remember to run it. You define a role once, and it keeps executing against new data in the background: new profiles that match, new signals on people already in your results, updated read on who's likely receptive to an outreach.
That's the whole distinction: not a new kind of intelligence, just persistence applied to search. The agent extracts structured parameters from your search (role, must-have skills, seniority, location), lets you set priorities and hard exclusions on top of that, and then keeps executing on a cadence instead of waiting for you to click search again. Results land back in the same project as your manual search and CRM imports, so reviewing them doesn't mean learning a second tool or checking a second inbox.
Concretely, setting one up takes three steps. First, you create the agent directly from a search you've already run, which means it inherits the criteria instead of asking you to redefine the role from scratch. Second, you calibrate what matters most: which requirements are non-negotiable (an IS/IS NOT rule, like "must have production Kubernetes experience") versus which are a preference the agent should weight rather than enforce. Third, you set a ceiling on how many results you want and where they should land, then launch it. From that point, the agent runs on its own schedule until you pause or edit it.
What runs on its own, and what doesn't
Inside Glozo's Sourcing Agent, the recurring work runs on three signals: a matching summary explaining why a candidate fits what you asked for, a market value estimate so you're not chasing someone outside budget, and an Open to Offers read on whether they're likely to answer outreach at all. The agent evaluates all three continuously and emails you when it has something worth reviewing. On a typical search, only a small fraction of matched candidates carry the Open to Offers signal, which is the whole point: the agent's job is narrowing 200 plausible profiles down to the handful actually worth a call, not handing you 200 profiles faster.
What it doesn't do: decide who's hired, reduce a person to a single number, or send outreach without your say-so. Today the agent runs in manual mode, meaning it hands you a refreshed shortlist and you decide what happens next. An automatic mode and a finer-grained "keep this one, drop that one" calibration step are both designed but not live yet, so any explanation of an agent that claims full autopilot sourcing, from first search to sent interview invite, is describing a future state, not a shipping product, whichever vendor is behind it.
That distinction matters beyond marketing accuracy. New York City's Local Law 144, the one US law that mandates an independent bias audit for automated hiring tools, explicitly stops at the applicant line: the city's own FAQ on the law says it does not cover scanning a resume bank or reaching out to potential candidates, only tools used to screen someone who has actually applied for a specific role. Sourcing sits outside that requirement by definition. California is the sharper edge to watch instead: under its 2025 FEHA regulations, an agent that performs recruitment or screening on an employer's behalf can itself be deemed an "employer," which is a reason to keep a recruiter reviewing and deciding rather than letting an agent act unsupervised, regardless of what any specific state currently requires.
An agent versus a one-time search
| Step | One-time search | Sourcing agent |
|---|---|---|
| Finding candidates | Returns what matches right now | Keeps checking for new matches as the market changes |
| Compensation fit | You filter manually each time | Applied automatically on every new result |
| Receptiveness | Not visible without re-running the search | Re-checked continuously; you're notified when it changes |
| Your effort after setup | Re-run the search yourself to stay current | Review what lands in your inbox |
| Final call on outreach | You | You |
The last row matters more than the other four combined. Nothing about running continuously changes who makes the outreach decision. An agent widens what you see and narrows it to what's worth seeing; it doesn't replace the judgment call at the end.
Where the DIY version hits a wall
A lot of recruiters have already built something like this themselves, with Claude Code, Claude Skills, or a custom GPT: a scaffolded folder of prompts that turns a job description into a search brief, or a saved skill that screens resumes against a rubric. That's a legitimate way to get repeatable structure into a messy workflow, and it costs nothing beyond the time to set it up.
The wall shows up at the data layer. A DIY agent built on a general-purpose model can operate on whatever it can reach: files you give it, an ATS export, whatever's on the public web. It has no live read on the talent market, no compensation model built from actual job postings, and no way to tell a self-reported "open to work" flag from a passive candidate who's actually receptive but hasn't said so anywhere. That's the specific gap a purpose-built sourcing agent is designed to close, by running on data the DIY setup structurally cannot reach rather than on a bigger prompt. Glozo is building toward an MCP server to make that handoff between a general-purpose assistant and its own market data more direct, though that's still in development with no committed ship date.
If you're weighing the two paths, how AI sourcing agents compare to a custom GPT goes deeper on that specific tradeoff, and how to choose an AI sourcing agent covers what to check before picking one. For the DIY build itself, building a recruiting agent with Claude Code walks through the setup and where it hands off. And if the question is which vendor's agent to actually use rather than what an agent is, our tested comparison of the best AI sourcing agents is the buyer's guide version of this same territory.
Before you commit to either path, see how an agent narrows a real search down to who's actually worth calling instead of taking the pitch on faith.
The takeaway
An AI recruiting agent's job is narrowing, not deciding. It takes the work of re-running a search and re-checking who's actually reachable off your plate, and it should stop exactly there. If a tool's marketing describes it making the call on who gets an interview, that's a claim about where the category is headed, not where it is today.
See it running on a real search. Turn one search into a Sourcing Agent and watch it narrow the field on its own, free while it's in manual mode.