16 agents run my business in the US | Egor Karpovich
No Bullsh!t, April 22, 2026, 100 min, 677 views. Egor Karpovich, co-founder and CEO of his travel company (a corporate travel management company and tour operator for the mid-market — companies with a travel budget of 1-10 million dollars a year), tells Sergey Kuznetsov how, in three months, he built a business around an autonomous AI agent based on Claude Opus that coordinates around sixteen specialized agents: sales, partnership, SMM, an analyst, and a lawyer.
| Published | |
|---|---|
| Length | 100 min |
| Views on YouTube | 677 |
| Language of the episode | Russian |
| Where to watch and listen | YouTube, Spotify, Telegram |
Date, length and view count according to data from the episode page on YouTube, as of the page update date shown in the footer.
In short
The journey started with outsourced development in 2012-2013, product pivots after the 2014 crisis, a coworking space in Minsk, and a move into corporate travel in 2019-2020. From there, Egor went from Zapier and Make (peaking at 150,000 operations a month) to vibe coding and fully autonomous agents with cron jobs, memory, MCP servers, and a human operator from South Africa working for 5 dollars an hour. They discuss specific mechanics: enrichment through Apify instead of Apollo and Clay, signals from reviews on Capterra and G2, partnerships with banks and HR platforms, a self-learning lawyer, a Google Workspace service account for reading email, the economics of Claude subscriptions versus Chinese models like Z.ai, and the main takeaway: don't spread yourself thin, and sell faster than you polish the product to perfection.
Key ideas
- Agent autonomy isn't built through code but through a cron schedule of sequential steps (checking, enriching, qualifying) plus memory retained between runs, so the system carries the task through to a result on its own, without constant human involvement.
- Targeted enrichment through custom Apify actors turned out to be cheaper and more reliable than bulk databases like Apollo or Clay, because the data is verified in real time instead of being pulled from a stale cache.
- Sales signals can be gathered not only from job changes but also from negative reviews on G2, Capterra, and TrustPilot: a first message that references a specific review sharply increases response rates compared to a cold connection request.
- Agents are given a role instead of a personal name (sales, partnership, lawyer, analyst) so there's no confusion about who's responsible for what, and so the agent's identity literally describes its function.
- A read-only Google Workspace service account lets the analyst agent monitor the company's shared inbox, prioritize emails, and restore context for another agent's request in seconds instead of waiting on a human.
- It makes sense to keep the expensive model (Claude Opus) only for the truly important tasks and move routine monitoring and parsing to cheaper models, because at a volume of hundreds of emails a day the cost difference becomes significant.
- The main business takeaway isn't about tools: don't spread yourself across new niches, and sell faster than you polish the product or the agent to perfection, because launching iteratively with a client saves more than upfront refinement does.
The full transcript
The episode is recorded in Russian, and the full transcript of the conversation is on the Russian page of this episode. English and Spanish subtitles for the strongest episodes are rolling out on YouTube in waves.
Open the transcript