An AI recruiting agent is software that carries out a defined stage of a search on its own — building the brief, sourcing, outreach, screening, scheduling, or tracking an accepted offer to joining — instead of waiting for someone to click through each step. The useful ones are narrow, and the safe ones cannot change anything without a human confirming it.
That distinction matters more than the label. Below is what each of the six we run actually does, and what none of them can do.
Agent vs automation
Automation follows a path you configured in advance: when status changes to X, send email Y. It is predictable and it is genuinely useful.
An agent picks its own steps toward a goal. Given "find people who can own a multi-region failover rebuild", it decides what to search, notices that it has three candidates with scale experience but none with migration ownership, and goes back for that specific gap. It can evaluate its own output and correct it.
That is more capable and more dangerous, which is why the gating below matters as much as the capability.
The six
Brief turns a thirty-minute hiring-manager call into a structured brief: must-haves, deal-breakers, target companies, comp band and the pitch. This is the highest-leverage agent, because everything downstream inherits its criteria. A vague brief produces a plausible, useless shortlist.
Search builds a search spec from the brief, works LinkedIn, Naukri, GitHub, publications, portfolios and our own talent graph at once, and screens every profile against explicit criteria. Then the part that matters: it critiques its own per-criterion coverage, notices which requirements are thinly evidenced across the shortlist, and re-sources specifically for those gaps over several rounds. Output is ranked profiles with evidence per criterion written as full sentences — not a match percentage, which tells a recruiter nothing they can act on.
One hard-won detail: scoring needs deterministic ceilings. Left to score freely, a model will rank a candidate highly on overall impression while quietly ignoring a must-have they clearly do not meet. Caps that hold down anything unproven or contradicted are what make the ranking trustworthy.
Outreach writes to the person rather than the profile and sequences across email, LinkedIn, WhatsApp and voice, picking the entry channel per person and escalating when one goes quiet. It enforces working hours in the candidate's timezone, daily and weekly caps with a warm-up ramp, suppression lists, and an automatic pause if bounces spike. Copy is approved by a human before it sends. It discloses that it is AI when asked, and it stops the moment someone replies so a person picks up the conversation.
Screen runs the structured screen against the client's rubric and returns a transcript, a scorecard and a written case with comp, notice and location confirmed. The scoring blueprint is generated from the role rather than pulled from a generic template, and thresholds are calibrated against who was actually hired.
Schedule coordinates panels across calendars, sends links and reminders, handles reschedules, and chases feedback after each debrief. Unglamorous, and it removes more recruiter hours per week than anything else on this list.
Close watches accepted offers through the notice period, flags counter-offer risk, and escalates when a candidate goes quiet. In India this is where searches die — sixty-plus days between acceptance and joining, during which the current employer gets to make their case.
Conducting all six is Pilot, an agent with 122 tools that turns a sentence into work across the others.
What none of them can do
They cannot persuade someone. A senior engineer weighing a move is deciding about scope, ownership, who they would report to, and what they are giving up. That is a conversation with a person who has had it before.
They cannot read a room. The pause before someone answers "how is the team doing?" is information no transcript captures.
They cannot hold a relationship through a notice period. Close flags the risk; a human makes the call.
Gating: the part to actually evaluate
If you are assessing any agentic recruiting system, ask what it is prevented from doing.
The pattern we use: every tool declares whether it mutates anything. Read-only actions run immediately. Anything that changes something is staged as a proposal a human confirms, supported actions can be undone for up to 24 hours, there is a company-wide kill switch, and every mutation is written to an append-only audit log. Per-skill autonomy is off | suggest | auto and defaults to suggest, so nothing acts on its own until someone deliberately enables it.
The test is not whether a vendor promises a human in the loop. It is whether the system can be made to act without one by flipping a setting. If it can, the promise is a policy, not a guarantee.
Where this leaves recruiters
The work the agents remove is list-building, first-touch sequencing, calendar coordination and note-taking. That is most of a recruiter's week and almost none of a recruiter's value.
What is left is the brief, the pitch, the judgment call on a borderline candidate, and the conversation that changes someone's mind. Agents make a specialist faster. They do not make a generalist into a specialist.