---
id: "lib-2026-11-05-why-we-added-a-nitpicker-to-our-team-of-ai-agents"
type: "library"
title: "Why we added a nitpicker to our team of AI agents (Зачем мы добавили в команду ИИ-агентов душнилу)"
lang: "RU"
org: "SIBUR"
source_type: "Practitioner post"
date: "2026-11-05"
week: "2026-W45"
url: "https://nosov.ai/library/lib-2026-11-05-why-we-added-a-nitpicker-to-our-team-of-ai-agents/"
original_url: "https://habr.com/ru/companies/sibur_official/articles/1103812/"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "SIBUR's product lead and its build partner Omnius.team describe a production multi-agent research system with a separate critic; concrete architecture and failure cases, but the one outcome figure is self-reported with no evaluation data."
summary_basis: "full-text"
original_minutes: "8"
tags: ["business transformation", "manufacturing r&d", "multi agent critic", "expert in the loop", "regulatory checks"]
---

# Why we added a nitpicker to our team of AI agents (Зачем мы добавили в команду ИИ-агентов душнилу)

Tier B (Useful, limited): SIBUR's product lead and its build partner Omnius.team describe a production multi-agent research system with a separate critic; concrete architecture and failure cases, but the one outcome figure is self-reported with no evaluation data.

Summary based on the full text.

## In short

A Habr post (5 November 2026) from Digital SIBUR, the Russian petrochemical group, by its senior key-projects manager, with technical notes from Omnius.team. SIBUR built a multi-agent system that assesses ideas for replacing materials with polymers across properties, regulation, market and economics. Early reports looked convincing but carried errors only experts could spot, so the team added a critic agent on Claude that challenges the researchers' arguments before a human sees the report. Assessment time fell from 88 working days to up to two weeks.

## Key numbers

- 88 working days to up to 2 weeks: average time to assess one material-substitution idea before versus with the system, including checks and approvals
- 16: steps in the pipeline: 14 AI stages and 2 deterministic phases
- ~5x: higher cost of the fallback search (Sonnet with Anthropic web search) versus Perplexity Sonar Pro
- 3: deep refactorings since the first version was tested in March 2026, plus two speed-ups

## What it reports

- 16 steps: 14 AI stages and two deterministic phases. Main engine: Gemini Pro for complex reasoning, Flash for simpler tasks. Critic: Claude, chosen because it checked well but would be too expensive for all the work.
- Search: Perplexity Sonar Pro; fallback to Sonnet with Anthropic web search, about 5 times more expensive and slower.
- Agents score their areas; the final verdict ("pilot only", "needs checking", "do not replace") is set by fixed rules.
- The critic runs once, with no open-ended debate. If the synthesis rejects the critic's grounds for refusal, it must state why. Experts launch re-research.
- Example: for multilayer polymer packaging the critic stressed delamination risk after sterilisation and asked for tests before and after; the recommendation became "pilot only".
- Failures that produced rules: medical packaging assessed against building codes (led to expert-set industry segments); ABS plastic read as anti-lock brakes (industry glossary); purchase price taken as total cost (a total-cost-of-ownership method); citing a repealed standard (a currency and applicability check).
- Speed work: parallel search calls within an agent round, two pairs of independent stages run in parallel, caps on call waits and search rounds, less page content loaded.
- A "do not replace" verdict is treated as a useful result.

## How they got it

A practitioner account of building and iterating the system with domain experts since March 2026. No evaluation set, accuracy figures or cost per run are given.

## Limits

- The 88-days-to-two-weeks figure is self-reported, with no breakdown of how it was measured.
- No data on how often the critic changed a recommendation or was itself wrong.
- Partly written by the external build partner.

## Why it matters for companies

The useful pattern is the division of labour: cheap models do the research, a stronger model is used only as a single-pass critic, and verdicts come from fixed rules rather than the model. The failure list (wrong industry, wrong acronym, price instead of lifetime cost, repealed norms) is a practical checklist for any research agent used for investment decisions. The system also needed ongoing expert input after launch, which is a running cost to budget for.

## Key passages in translation

> Хуже всего то, что по самому ответу ошибку не заметишь: он подробный, логика, на первый взгляд, стройная. Вот только решение по нему принимать нельзя.

The worst part is that you cannot spot the error from the answer itself: it is detailed and the logic looks sound at first glance. You just cannot make a decision on it.

> Раньше оценка идеи занимала в среднем 88 рабочих дней [...] Теперь с системой — до двух недель вместе с проверками и согласованиями.

Assessing an idea used to take 88 working days on average [...] Now, with the system, it takes up to two weeks including checks and approvals.

> Критик, увы, тоже может ошибаться, поэтому живой эксперт проверяет и исходные аргументы, и замечания к ним.

The critic, alas, can also be wrong, so a human expert checks both the original arguments and the objections to them.

Original: https://habr.com/ru/companies/sibur_official/articles/1103812/