---
id: "lib-2026-10-20-code-gets-cheaper-verification-gets-dearer"
type: "library"
title: "Code gets cheaper, verification gets dearer"
title_original: "ИИ в разработке: код дешевеет, проверка дорожает"
lang: "RU"
org: "Гриненко про (Manychat)"
source_type: "Podcast"
date: "2026-10-20"
week: "2026-W43"
url: "https://nosov.ai/library/lib-2026-10-20-code-gets-cheaper-verification-gets-dearer/"
original_url: "https://www.youtube.com/watch?v=9d5JXwNeVjI"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "First-hand account from a long-serving Manychat engineering leader on hiring, team shape and verification under coding agents; candid and specific, but opinion-led with few hard numbers."
summary_basis: "transcript"
original_minutes: "88"
tags: ["team design", "hiring freeze", "token budgets", "technical interviews"]
---

# Code gets cheaper, verification gets dearer

Original title: ИИ в разработке: код дешевеет, проверка дорожает

Tier B (Useful, limited): First-hand account from a long-serving Manychat engineering leader on hiring, team shape and verification under coding agents; candid and specific, but opinion-led with few hard numbers.

Summary based on the full transcript.

## In short

An 88-minute Russian podcast episode (Grinenko Pro, uploaded 20 October 2026) with Sergey Sergeev, who has worked at Manychat, a chat-automation SaaS, for about 6.5 years and now applies AI to business and operations processes. His thesis: code generation became cheap this year, and the cost moved to verification, CI and trust. Manychat paused hiring at its former volume because nobody knows yet which team configuration works when agents act as extra developers. Interviews no longer distinguish candidates.

## Key numbers

- 4 teams of 2: team configuration the guest runs while not hiring
- two two-week sprints: time most teams spent entirely on learning agent tools and experimenting
- ~EUR 200: starting tier of the internal AI subscription budget per person, raised on request
- 250,000: illustrative count of passing generated tests that may check nothing if nobody reviews what they test

## What it reports

- Git still fits agent work; commit messages matter more because nobody reads diffs of -1,000/+2,000 lines. CI became a bottleneck as iterations got cheaper and more frequent.
- Teams handling financial and personal data still read code; frontend teams read much less. Manychat has never had a QA function; developers test their own code.
- Generated tests are cheap and can be empty: a model can report that 250,000 tests passed while testing nothing. His rule: design how you would verify a feature written by AI without looking inside it, and isolate agents from personal data.
- Hiring paused at former volume "not because we do not need hands" but because effective configurations with virtual developers are unclear.
- The guest runs "virtually four teams of two people" and is not hiring. Specialisation is dissolving: each person ends up carrying the whole system in a two-person team.
- Most teams took about two two-week sprints entirely for learning tools and experimenting.
- Token budgets: internal subscription tiers starting around EUR 200, raised on request; nobody has been told off for spending.
- Interviews: he stopped giving coding tasks; architecture questions are also easy to answer with AI help; a probation period no longer reveals long-term fit.
- Replicating Manychat's product is "probably not very hard"; the moat is product vision, marketing money and distribution.

## How they got it

A recorded conversation. The guest describes his own teams and company practice; he labels several views as his own and possibly overhyped. The digest is based on a cleaned auto-caption transcript.

## Limits

- Self-reported experience from one company; no metrics on output, quality or cost.
- Auto-captions may garble names and figures.
- The guest now works on operations rather than product engineering and says so for some topics.

## Why it matters for companies

The concrete signal is the hiring pause: a mature SaaS firm is holding headcount until it knows how to combine people and agents, not because work disappeared. Budgeting a two-sprint learning break and open token tiers is a cheap, copyable enablement pattern. The verification point applies to any function adopting agents: define acceptance checks that do not depend on reading the output line by line, and keep agents away from personal data. Interview design for technical roles needs rework, since standard coding and architecture questions no longer separate candidates.

## Key passages in translation

> мы приостановили наём в том объёме, в котором он был раньше, потому что это не не не потому, что нам не нужны руки, а потому что непонятно, как в каких конфигурациях сделать эффективную работу.

We paused hiring at the volume it was before, not because we do not need hands, but because it is unclear in which configurations to make the work effective.

> Поэтому у меня сейчас э-э назовём виртуально четыре команды по два человека, и я пока не нанимаю.

So right now I have, let's call it, virtually four teams of two people, and for now I am not hiring.

> если не проверять, что она тестирует, она просто тебе скажет, что 250 тыс. тестов прошли успешно, погнали в продакшн, а что там, непонятно.

If you do not check what it is testing, it will just tell you that 250,000 tests passed, off to production, and what is in there is unclear.

Original: https://www.youtube.com/watch?v=9d5JXwNeVjI