Hiring handed to agents
KARPOV.CONF 2026: найм и развитие команды отданы ИИ-агентам
B · Useful, limited A founder (ex-Stepik, CEO of Hyperskill) describing a step-by-step recruiting pipeline with before-and-after time figures, but at small scale, self-reported and presented on his own agent platform, so it doubles as a product pitch.
Summary based on the full transcript·Original: about 27 minutes to watch
In short
A 27-minute talk at KARPOV.CONF 2026 (28 October 2026) by Ruslan Davleshin (surname as given by the host, unverified), formerly of Stepik, now CEO of Hyperskill and founder of an agent-workspace startup whose name is garbled in the captions. He describes two internal cases built on his own platform: a hiring funnel where agents score CVs, chase missing information, check inflated CVs, prepare questions and write interview reports; and a team-development set-up where agents keep a running file on each staff member from meeting transcripts and chat. He frames the work as three levels: personal automation, agents for other people, and agents that run inside other people's agents.
Key numbers
- 150-200applications a week for engineering and product-engineering roles, against 40-60 a recruiter could review by hand
- 18 h to ~3 hrecruiter time across all stages (CV review 5 to 0, preparation 6 to about 2, interview processing 7 to 1)
- 25 to 10 daystime from entering the funnel to an offer
- 1 h vs 2-3 daystime to get an interview report after each stage
As reported by the source.
Russian·Video·opens youtube.com
What it reports
- Baseline: 150-200 applications a week; a recruiter could review about 40-60 by hand, so strong candidates were lost and there was no time for new sourcing channels or employer brand.
- CV scoring: an agent scores each CV 0-100 against role expectations, using the team's prior assessment know-how.
- A sub-agent emails candidates whose CVs are thin, usually two rounds of three questions. This forced a change of applicant tracking system, because the old one's API did not return card comments.
- A second sub-agent checks CVs over-tailored with AI by looking at LinkedIn history, running web research and following project links.
- For each interview stage (screening, case, technical, final) agents generate questions, analyse the transcript and send a report; the interviewer accepts or comments, and a sub-agent re-scores the candidate. The final report proposes a grade and a compensation range; the founder and recruiter decide.
- A "founder" agent trained on the speaker's own interview transcripts feeds his criteria into the recruiter's question and scoring agents.
- The recruiter built and maintains all hiring agents without developers, using the platform's agent architect.
- Results: all applications are reviewed and ranked; interview reports arrive within an hour instead of two to three days; recruiter time fell from about 18 hours to about 3 hours; time to offer fell from about 25 to 10 days. Freed time went to new sourcing channels and the employer brand.
- Team development: three agents process 1:1 transcripts into a per-person file, pull positive peer feedback and value-aligned behaviour from Slack, the tracker and team meetings, and draft 1:1 agendas. An HR-built 360 agent sets goals from the same file. Results are qualitative only.
- Level-three examples: a product-manager agent and SOC 2 compliance agents that feed context into Claude Code and Codex through MCP.
- Q&A: on privacy, he said the team quickly got used to all meetings being recorded. For clients bound by Russian data law (152-FZ) the system runs on-premises with local models.
How they got it
The speaker reports on his own company's use of his own product. Time figures are his estimates of recruiter hours per stage before and after; no hiring-quality measure (offer acceptance, retention, performance) is given.
Limits
- Self-reported by a vendor about its own platform; the cases are presented partly as a product demonstration.
- Small scale: one recruiter and a weekly flow of 150-200 applications. The 18-hour figure is not stated as per week or per hiring cycle.
- No data on scoring accuracy, false rejections or bias in CV scoring, which are the main risks in automated screening.
- The privacy answer rests on staff getting used to recording; no consent or data-retention rules are described.
- Source is YouTube auto-captions: the startup name, the speaker's surname and several numbers ("150 тире2", "4060") are garbled and reconstructed.
Why it matters for companies
Recruiting is one of the first HR processes firms try to automate, and this talk is a concrete map of where the time goes: CV triage, chasing missing information, writing interview feedback. The fastest gain came from reports written straight after each interview, which removed days of waiting on engineers' feedback. The ATS change is a practical warning that agents need API access to everything the process touches. The employee-file case shows the trade-off clearly: continuous recording of 1:1s and chat gives agents context, and raises consent and trust questions that need explicit rules, plus a check of automated screening for bias.
Key passages in translation
Russian original
А, и к нам примерно в неделю приходит 150 тире2 заявок. А, и до появления, ну, и я рекультёр руками успевал разбирать примерно 4060.
We get about 150 to 200 applications a week, and before this a recruiter could review about 40 to 60 by hand.
Russian original
Раньше у нас было где-то 25 дней. ... Сейчас это сократилось за 10 дней благодаря э использованию идеагентов.
Before, it took about 25 days. Now it is down to 10 days thanks to the agents.
Russian original
У нас получилось так, что мы как будто бы достаточно быстро привыкли к тому, что все встречи всегда записываются.
It turned out we got used fairly quickly to all meetings always being recorded.
Summary based on the full transcript. Read the original for the full argument and data.Watch the original on youtube.com
- Tags
- Future of work