Can a robot write a merge request? Frontend agents
Может ли робот написать merge request? Агенты фронтенда с GigaChat и без
B · Useful, limited SberDevices frontend-infrastructure engineer's company-blog case on in-house review, log-analysis and build-fix agents; volumes and comment categories are given, but no acceptance rate, fix success rate or comparison with a baseline.
Summary based on the full text·Original: about 9 minutes to read
In short
A Habr post (13 October 2026) on the SberDevices company blog by Avrora, who works on frontend infrastructure. External tools such as GitHub Copilot Code Review were ruled out because the team did not want to expose its code, so it built its own set of frontend agents on GigaChat. They mix deterministic workflows, where code drives the model through narrow steps, with one true agent loop for fixing broken builds. The post covers the AI reviewer's findings, a log parser that narrows context, and the feature people used most, which turned out to be an automatic call for testers.
Key numbers
- 4,000+merge requests checked by the GigaChat-2-Max reviewer across 90 projects, January to July 2026
- 8%share of those merge requests where the reviewer flagged a potential problem, about one in 13
- 18% / 16% / 12%share of reviewer comments on logic errors, null/undefined and error handling, and types/API contracts
- almost 4,000reminders sent by the review-nudge bot in 51 projects, January to August; median 2 per merge request
- 10 to 30calls a day to the log-analysis and build-fix agent, at least 1,800 in total
As reported by the source.
Russian·Web page·opens habr.com
What it reports
- The reviewer is GigaChat-2-Max with no fine-tuning, run as a fixed workflow: GitLab webhook, diff fetch, file triage into "needs review" or "approved", model call at temperature 0, LGTM filtered out, line-level comments plus an MR summary.
- Over 4,000 MRs in 90 projects reviewed January to July; problems flagged in 8%, usually one or two comments each. The author attributes the low rate to developers already running local agents before review.
- Over 90% of comments point to a concrete problem. Topics: style and naming 20%, logic 18%, null/undefined and error handling 16%, types and contracts 12%, syntax and build 11%, the rest 23%.
- Log analysis: a parser keeps only the last failed command's output, capped at 10,000 characters, strips secrets, and asks GigaChat to say whose side the problem is on and what to do.
- Build fixing: a full agent with file, search, edit and shell tools; a human must press "Try fix with AI"; if CI passes, it opens an MR that a human and the AI reviewer still approve. Supported for app builds, Docker builds and nginx config.
- Two human approvals remain mandatory; all AI comments are advisory.
How they got it
Counts from the team's own systems (GitLab, MongoDB, PostgreSQL logs of prompts, answers and like/dislike reactions). The comment categories come from the author's own classification of 2026 reviewer comments.
Limits
- No acceptance rate for reviewer comments and no success rate for automatic build fixes.
- No comparison of projects with and without the reviewer or the reminder bot; the author notes the bot is on by default, so there is no control group.
- Company blog post; GigaChat is the parent group's own model.
Why it matters for companies
The design point is that most of the value came from narrow, code-driven workflows around a model, with a free-running agent reserved for one task that has an objective check: does the build pass. The context parser is a cheap, general lesson: send the model the last failed command, not the whole log. The author's closing view matches other reports this week: code generation is getting cheaper faster than verification, so tests, types, reproducible builds and simple coordination tools such as reviewer reminders deserve the investment.
Key passages in translation
Russian original
Технически AI-ревьюер — это не AI-агент. Его workflow задан кодом
Technically the AI reviewer is not an AI agent. Its workflow is defined in code.
Russian original
С развитием искусственного интеллекта генерация кода быстро дешевеет — быстрее, чем проверка.
As AI develops, code generation is getting cheaper fast, faster than checking it.
Russian original
самой востребованной функцией системы агентов оказался… автоматический призыв тестировщиков в проект после обновления кода.
The most requested function of the agent system turned out to be automatically summoning testers to the project after a code update.
Summary based on the full text. Read the original for the full argument and data.Read the original on habr.com
- Tags
- Coding agents