How development and team setup change with agents
Как меняется разработка и сетап команд с внедрением агентов
B · Useful, limited Avito's head of platform development gives a rare company-level throughput metric and an honest account of where agent gains leak away; the figures are self-reported from slides not shown in the transcript.
Summary based on the full transcript·Original: about 41 minutes to watch
In short
A 41-minute conference talk (uploaded 20 October 2026 on AvitoTech) by Alexander, head of platform development at Avito, the Russian classifieds marketplace, who leads its agent-based SDLC programme. Avito spent 2025 building its own assistants with weak adoption, then in February 2026 switched to off-the-shelf agent harnesses and put the effort into context: integrations, skills and secure access. It dropped "minutes saved" metrics and now tracks significant production increments per constant headcount. The gain is real but shrinks to single-digit business impact, and review time fell because people look less.
Key numbers
- -15%2026 cycle-time target, already beaten by July
- 30%+rise in significant production increments, while all pull requests nearly doubled
- 100+ / 4,000+MCP servers in Avito's auto-discovery hub and skills in its skills hub
- ~5parallel local agent sessions per person before cognitive limits stop scaling
As reported by the source.
Russian·Video·opens youtube.com
What it reports
- February 2025: in-house tools and fine-tuning on Avito's code base; adoption was the key metric and "went quite badly". June 2025: open-weight coding models became good enough for agent pipelines. February 2026: stopped building own assistants, adopted top harnesses, focused on context engineering for all functions.
- Time-saving metrics were dropped: an hour saved per developer per day gives "no business effect by default" because it dissolves.
- New metrics: an internal benchmark of agents on frequent Avito tasks, and company-wide cycle time as significant increments per headcount. Target -15% for 2026, beaten by July.
- Enablers: pre-configured agents, anonymisation and a secure perimeter for external models, an MCP hub with 100+ servers and a skills hub with 4,000+ skills judged against golden sets.
- Significant increments rose more than 30% and all PRs nearly doubled, yet review time fell. Multiplied by the normal success rate of product changes, the business gain is only a few percent.
- Bottlenecks named: code review, product discovery, communication. A typical change passes 5-12 roles; teams where one engineer delivered backend, frontend and mobile end to end gained much more than functional teams.
- Non-engineers build MVPs in 1-2 sprints in isolated environments with no internet, then test them on real traffic under A/B tests; most of that code is later rebuilt.
- Local agent work stops scaling at about five parallel sessions per person; the next step is autonomous end-to-end pipelines with humans reviewing finished, spec-compliant PRs.
How they got it
Internal measurement presented at a conference. The speaker describes the metric definitions but the charts are on slides that the transcript does not capture. The digest is based on cleaned auto-captions in which product names and the speaker's surname are garbled.
Limits
- Self-reported company figures; no baseline values, confidence intervals or method for "significant increment" shown.
- The -15% and 30% figures come without absolute numbers.
- The speaker's surname could not be verified from the captions.
Why it matters for companies
The metric choice is the transferable part: Avito abandoned summed minutes saved because they never showed up in output, and measured delivered increments per constant team instead. The talk also shows where gains leak: faster review that is really less review, and many hand-offs between roles. That supports smaller, cross-platform teams and owner accountability for what ships. Leaders setting AI targets for engineering can copy the output metric and the honesty about single-digit business effect.
Key passages in translation
Russian original
если посмотреть на бизнес-эффект, это не даёт абсолютно никакого бизнес-эффекта по умолчанию.
If you look at business effect, it gives absolutely no business effect by default.
Russian original
ревью проходит быстрее просто, ну, путём того, что люди меньше просто, банально меньше смотрят. изменения.
Review goes faster simply because people, plainly, look at the changes less.
Russian original
помножить, да, эти две вещи друг на друга, то там в итоге это на самом деле всего проценты, единица процентов выигрыша.
Multiply these two things together and in the end it is really just a few percent, a single-digit percent gain.
Summary based on the full transcript. Read the original for the full argument and data.Watch the original on youtube.com