An entrepreneur replaced almost all staff with agents: 10 rules
Предприниматель уволил почти всех сотрудников и заменил их ИИ-агентами. 10 правил
C · Weak signal A solo talk by Slava Sazhin, co-founder of an agent-deployment firm, about his other company; the control architecture is described in detail, but every outcome is unaudited, self-reported marketing, including a claim of zero errors.
Summary based on the full transcript·Original: about 32 minutes to watch
In short
A 32-minute Russian talk (uploaded 21 October 2026) on the Combobulating channel, a firm that sells AI agent deployment. Co-founder Slava Sazhin describes his other company, which serves doctors, pharmacists and pharma clients. It had 100 staff five years ago, 50 a year ago and 25 in March 2026; in mid-April he cut to six. He keeps people required by law and people who meet clients in person, and runs the rest with Claude Code and Codex agents. The talk is framed as ten rules, and its core is spending most tokens on verification.
Key numbers
- 25 to 6staff at the speaker's company in March 2026 and after mid-April 2026
- 20% / 80%share of tokens spent on doing work versus checking it
- 4xtokens used by sceptic agents relative to the working agents
- ~70hooks that stop a model when it skips a mandated sub-agent
- 3 a month to 6 a dayclient presentations before (two people) and after (agents), as claimed
As reported by the source.
Russian·Video·opens youtube.com
What it reports
- Remaining six: CEO and chief accountant (legally required) and four people who meet clients face to face. Everyone "sitting at a computer typing" left.
- Rules: use only Claude Code or Codex, never chat; only the top model at the highest effort setting.
- Architecture: hundreds of scheduled agents write to a shared roadmap document; four read it ("repair", "build", a nightly read-only "owner" agent, and a "CEO" agent running about every 15 minutes). Collectors do not decide; deciders do not execute. He is asked about three times a day.
- Controls: thousands of rule files, about 70 hooks, watchdog agents, weekly red-team agents, A/B tests on every customer-facing change.
- Verification: 20% of tokens on work, 80% on checking; sceptic agents use four times the tokens of working agents; at least two sceptics per task, up to eight rejection rounds; a three-judge panel issues a 15-second password that a hook requires before any action ships.
- Permissions: the main agent can access his correspondence, bank accounts and SMS; hard bans only on sending his personal money and signing documents.
- Output claim: presentations went from three a month by two people to six a day, allowing about 30 tenders a month instead of three.
- Recipe for replacing a role: record a staff member's screen for a day and have a coding agent build an agent from it.
How they got it
A monologue describing the speaker's own company. No financials, error logs, quality data or customer outcomes are shown. The digest is based on cleaned auto-captions.
Limits
- Self-reported and unaudited; the channel sells agent deployment.
- The claim that the system "has not made a single error" is implausible without logs and is not evidenced.
- Granting an agent access to bank accounts and all correspondence is a security posture most firms could not accept.
- No data on what happened to quality, clients or revenue after the cuts.
Why it matters for companies
Treat the headcount story as marketing. The useful content is the control pattern: separate collection, decision and execution; spend most tokens on independent checking; use hooks and short-lived approval tokens so no action ships without a vote. That pattern transfers to any agent workflow with real-world side effects. The role-replacement recipe (record the work, build an agent from it) also shows how tacit process knowledge is captured, and why staff may resist screen recording.
Key passages in translation
Russian original
из 100% токенов, которые я трачу, 20% идёт на работу и 80% на её перепроверку.
Of 100% of the tokens I spend, 20% go to the work and 80% to re-checking it.
Russian original
У меня на скептиков уходит в четыре раза больше токенов, чем на самих агентов.
My sceptics use four times more tokens than the agents themselves.
Russian original
Все те, кто сидят на попе ровно за компьютером, что-то там ручками печатают, мне не нужны.
All those who sit still at a computer typing something by hand, I do not need.
Summary based on the full transcript. Read the original for the full argument and data.Watch the original on youtube.com