Monday, 9:04 am: a recruiter opens a new req

Priya has been recruiting for eleven years. She remembers when a new requisition meant a blank search bar. You would take the job description, guess at keywords, and start scrolling through LinkedIn until your eyes glazed. A senior data engineer req might mean four hundred profiles viewed to find sixty worth a second look, and maybe twenty worth a message. That was a week of work, and it was the part of the job she liked least.

This Monday looks different. The req for a senior data engineer went live Friday afternoon. Priya approved the requirements, closed her laptop, and went home.

Over the weekend, a sourcing agent went to work. It read the job description, pulled the must-haves apart from the nice-to-haves, and started searching. Not keyword matching. It read profiles the way a person would, if a person had infinite patience: this candidate says "built streaming pipelines" but the dates suggest she inherited them; this one lists Spark but his last two roles were analytics, not engineering; this one never mentions the exact stack but shipped something structurally identical at a logistics company. By Sunday night it had looked at just over two thousand profiles and kept two hundred and forty.

Then a second agent took the handoff. This one had a narrower job: score the two hundred and forty against the actual requirements and rank them, with reasoning attached to every rank. Not a number. A paragraph. "Ranked 4th: strong pipeline experience at comparable scale, gap in orchestration tooling the team uses, but her open source work suggests she picks up tooling fast. Risk: two short tenures in a row."

Priya opens the hiring channel at 9:04 and the shortlist is sitting there, posted at 6:12 am, forty names with reasoning, waiting for a human.

The part people miss

When teams describe this setup to me, the reaction from people who haven't seen it is usually some version of "so the machine picks the candidates." That is exactly what is not happening, and the distinction matters more than the speed.

The machine ranked the candidates. Ranking is an opinion with evidence attached. Picking is a decision, and the decision still belongs to Priya. She spends her Monday morning doing what she is actually good at: reading the reasoning, disagreeing with it, and saying so.

She bumps a candidate from 14th to her top five because the reasoning undervalued a career gap that Priya reads differently. She kills two of the top ten because she has phone-screened both of them before at a previous company and knows things no profile will ever say. When she overrides a rank, she writes a sentence about why, and the ranking agent takes it in. Not as a correction to a bug. As information about how this team weighs things, which shapes how the next shortlist gets built.

This is the texture of the change. It is not a tool spitting out a list. It is closer to working with a very fast junior sourcer who never sleeps, always shows their reasoning, and never gets defensive when overruled.

Where the agents talk to each other

Here is the part that would have sounded like science fiction in 2023 and is now just how certain teams run.

The hiring manager, Tomás, has his own agent watching the req. When Priya finalizes her top fifteen around 10:30, she doesn't email Tomás a spreadsheet. Her shortlist lands where his agent can see it, and his agent does its own pass before he ever looks: cross-checking candidates against the team's actual codebase and current skill gaps, not the job description's idealized version of them. It flags two candidates as stronger fits than their rank suggests, because the team's real bottleneck is a data quality problem the JD barely mentions.

The two assessments don't match. The sourcing side ranked for the req as written; the hiring manager's side ranked for the team as it exists. So the agents do something that still catches people off guard the first time they watch it happen: they reconcile. One posts its reasoning, the other posts its counter, and where they can't agree, the disagreement gets surfaced to the humans as exactly that. A disagreement, with both arguments laid out.

Tomás reads it over lunch and settles it in about four minutes. He knows things neither agent can know, like the fact that the data quality problem is getting a dedicated hire next quarter. But he settles it with both cases in front of him, instead of settling it the old way, which was not settling it at all because nobody had time to construct either argument.

The two hundred who didn't make it

One more detail from that Monday, easy to miss and possibly the most valuable thing in the whole setup.

The two hundred candidates who didn't make the shortlist don't vanish. Each of them keeps their reasoning: close but light on distributed systems, strong but almost certainly out of the compensation band, excellent but wrong seniority for this req. In the old world, that near-miss knowledge lived in a recruiter's head for about a week and then was gone forever. Every new req started the archaeology from scratch, and the same promising person got discovered, evaluated, and forgotten by the same company three separate times over two years.

Now, when a staff-level version of this role opens in the fall, the sourcing agent starts by rereading its own graveyard. Four of the spring's near-misses are suddenly dead-center fits, and the outreach to them can honestly say: we looked at you carefully six months ago, here's what's different now. Recruiters have talked about "building a talent pool" for decades. Mostly it meant a spreadsheet nobody maintained. A pool that maintains itself, remembers why each person is in it, and rereads itself every time a req opens is a different animal.

What actually changed

The honest accounting is this. Priya's Monday used to be forty hours of profile scrolling spread across a week and a half. Now it is three hours of judgment on Monday morning. The req that used to take twelve days to produce a shortlist produced one over a weekend.

But the more interesting change is qualitative. Every candidate on that list has a written case for and, where relevant, against. When a candidate asks for feedback after a rejection, there is something real to draw on. When the team wonders in March why they keep losing candidates from a certain source, the reasoning trail from January is still there to interrogate.

Recruiting has always run on judgment, and the judgment was always trapped in people's heads, evaporating the moment a req closed. The thing these teams have now is judgment that leaves a residue.

There are failure modes, and the good teams are blunt about them. A ranking agent trained on your past hires will happily reproduce your past biases with a confidence score attached, which is why the reasoning-attached-to-every-rank practice exists in the first place. You cannot audit a number. You can audit a paragraph. The teams doing this well treat every ranking as a claim to be challenged, and they staff the challenging. The teams doing it badly treat the ranking as an answer, and they will spend 2027 discovering what they optimized for.

By 11:40, Priya has approved outreach to twelve candidates. The drafting of those messages is another handoff to another agent, and another story. She gets up to make coffee, and on her screen the hiring channel scrolls quietly on its own: the sourcing agent confirming the twelve, Tomás's agent updating the loop plan, a thread growing under a req that is four days old and already further along than last year's version got in three weeks.

The blank search bar is still there, somewhere, in a tab she hasn't opened since Friday.