Recruitment Agent Stack

The problem this solves

A typical recruiter at a fast-growing fintech is averaging about 1.3 hires a month. The bottleneck isn't candidate supply — it's the volume of low-leverage human attention required at the top of the funnel. Profile review. First calibration calls. Scheduling. Chasing candidates who've gone quiet. The strategic 30% of the role — calibrating with hiring managers, advocating for candidates, closing offers — gets squeezed.

The Recruitment Agent Stack flips that ratio. The agents do the boring 70%. Recruiters do the strategic 70%.

The four agents

1. Sourcing match agent. Takes a job spec and a candidate pool. Scores candidates against the actual role (not keyword matches), explains why the top candidates fit, and flags edge cases the recruiter should look at personally.
2. Voice screening agent. Runs a 10-minute calibration call — open / not open, salary expectation, work authorisation, basic competency check. Writes a clean summary the recruiter reads in 90 seconds. Candidates report it as fine; sometimes better than the human version (no judgement, no awkward small talk).
3. Scheduling agent. Owns the "does Tuesday at 2 work for both of you" loop. Reclaims ~4 hours per recruiter per week.
4. Pipeline hygiene agent. Nudges candidates who've gone quiet, flags stalled requisitions, writes the weekly pipeline summary nobody had time to write before.

How it actually works

Each agent is a small, focused thing — not a single mega-prompt. They're stitched together in Python with the Claude API as the reasoning layer, calendar APIs for scheduling, a voice model for the calls, and direct integration into the ATS so the work product lands where recruiters already live.

The interesting design choices are mostly about what NOT to automate. We deliberately leave hiring-manager calibration, candidate advocacy, and offer negotiation entirely human. Those are the moments that make hiring good or bad. Automating them would be a downgrade.

What's hard

• Calibration with hiring managers. Agents are only as good as the brief. A vague job description gets vague output, faster. The discipline of pinning down "good" is now the entire game.
• Bias compounds at speed. An agent calibrated on last year's hires reproduces last year's blind spots — faster. Needs explicit, ongoing attention. Not a one-time audit.
• Candidate experience. Agents are quick. Quick can feel transactional. The human moments matter disproportionately when 80% of the funnel is automated.

"By 2027, 'manages AI agents' will be a line on most senior ops job descriptions. The companies that win will be the ones that learned to do this two years before their competitors."

What's next

v2 brings two things: better feedback loops (so the sourcing agent learns from which candidates actually convert through interview rounds), and a hiring-manager-facing agent (a calibration assistant that helps managers write better briefs in the first place). Rolling both out across H1 2026.

Want to build something similar?

I'm happy to share what worked, what didn't, and which vendors are doing real work vs. which are theatre. The patterns transfer to most other operational functions — recruitment is just where the gains are most measurable.