SIBUR added an AI critic to its multi-agent system and cut idea evaluation from 88 days to two weeks

Every team needs someone who asks at the right moment: "Where did these numbers come from?" At SIBUR, the company decided to turn that skeptic into a dedicated AI agent, and it proved to be one of the project's most useful design choices.

The problem they were solving

SIBUR produces polymer feedstock, which its customers use to make packaging, pipes, syringes, and windows. To expand its market, the company continually evaluates substitution ideas: whether a component currently made of metal could instead be made from a polymer. Each idea requires extensive research into material properties, operating conditions, regulations, demand, and economics. Previously, evaluating a single hypothesis took an average of 88 working days. With the multi-agent system, the process takes up to two weeks, including reviews and approvals.

How the system works

The system has 16 steps: 14 AI stages and two deterministic phases. Research agents collect arguments across different areas, while the critic checks whether those arguments hold up, whether any constraints have been overlooked, and whether there are contradictions. The main engine runs on Google Gemini: Pro handles complex reasoning, while Flash handles simple tasks. Claude was selected as the critic because it is good at verification, although running the entire workflow on it would be too expensive. Perplexity Sonar Pro is used to find context, while the fallback option, Sonnet with Web Search, is slower and roughly five times more expensive.

Here is an example of the critic at work: a researcher writes that a material is cheaper and therefore the substitution is cost-effective. The critic checks exactly what was compared. If it was only the purchase price, that is far from enough to draw a conclusion about total cost of ownership. In another run involving multilayer packaging, it amplified the risk of delamination after sterilization and required testing before and after processing. The recommendation received a "pilot only" status.

There is no endless debate: criticism is a single stage. An expert initiates any repeat research, and if the final synthesis rejects the critic's argument, it must explain why. Agents assign the scores, while the verdict is determined by predefined rules.

Where the system broke down

The initial answers looked convincing, but experts found errors. Detailed, coherent prose can conceal a faulty premise.

• The system evaluated medical packaging using construction SNiP standards. The solution was segment-specific configuration: an expert answers a set of questions once and uploads documents, which colleagues can then reuse.
• "ABS plastic" turned into "brakes." An industry glossary was introduced.
• The system compared purchase prices while overlooking operating costs. A total-cost-of-ownership methodology was added, and missing data are now made visible in the result.
• An agent cited a repealed standard. The system now checks whether a document is current and applicable.

Another problem involved the word "obechaika," meaning the sleeve around a cup: no matching segment existed in the catalog. Users therefore had to be given the ability to add their own segments.

How it became a practical tool

Speed: independent search calls and two pairs of stages—regulations with economics, and market analysis with consumer properties—run in parallel. The team limited the number of search rounds and the volume of downloaded pages, and removed duplicated context from the final synthesis. Value: "do not substitute" counts as a full-fledged result because it can save months. Stability: new segments, documents, and model changes require continuous maintenance.

Key takeaways

If you are building an agentic system for decisions where errors are costly, make verification a separate role and, where possible, run it on a different model. Define verdict rules outside the LLM, add a domain glossary, and verify that sources are current. Keep a human at the end of the process: the critic can also be wrong. One limitation of the material is that it provides no figures on the critic's quality, so it is unclear how many errors the agent actually catches. Still, the idea of separating "generation" from "challenge" transfers well to other domains.