---
id: "lib-2026-11-02-don-t-wait-for-1c-updates-you-have-deepseek-1"
type: "library"
title: "Don't wait for 1C updates. You have DeepSeek (Не ждите обновлений от 1С. У вас есть DeepSeek)"
lang: "RU"
org: "Habr-Author"
source_type: "Practitioner post"
date: "2026-11-02"
week: "2026-W45"
url: "https://nosov.ai/library/lib-2026-11-02-don-t-wait-for-1c-updates-you-have-deepseek-1/"
original_url: "https://habr.com/ru/articles/1102948/"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "Evgeny Sokolov's Habr post runs a blind test with a real answer key (1C's own official patches) and publishes the sessions, cost and time per case, but covers only three bugs and one feature request, scored by the author."
summary_basis: "full-text"
original_minutes: "8"
tags: ["ai technology", "enablement", "1c erp maintenance", "coding agents", "cost per fix"]
---

# Don't wait for 1C updates. You have DeepSeek (Не ждите обновлений от 1С. У вас есть DeepSeek)

Tier B (Useful, limited): Evgeny Sokolov's Habr post runs a blind test with a real answer key (1C's own official patches) and publishes the sessions, cost and time per case, but covers only three bugs and one feature request, scored by the author.

Summary based on the full text.

## In short

A Habr post (2 November 2026) by Evgeny Sokolov testing whether a cheap model can fix bugs in a million-line 1C ERP configuration as well as 1C's own developers. The agent (DeepSeek V4.1-Flash) got only a one-line, user-level symptom from 1C's bug list for UNF 3.0.12; its fix was then compared with the official patch held back as the answer key. All three bugs were fixed through configuration extensions, without touching the standard configuration. A larger feature from release 3.0.14 worked but was architecturally worse than 1C's version.

## Key numbers

- 155-427 RUB ($1.93-5.30): DeepSeek API cost per bug fix across the three bug cases
- 25-47 minutes (160-232 steps): time and agent steps per bug fix
- 2 of 3: bug fixes that hit the same code location as 1C's official patch (one an exact match); the third worked but was less precise than the vendor's
- 2 h 1 min, 747 RUB ($9.28): time and cost of the multi-step feature build with UI testing (case 4)

## What it reports

- Case 1, "division by zero" when printing a UPD for quantities below 0.006: the agent found the same query line 1C patched in a 12,000-line module, added a guard on the rounded divisor, and noticed 1C's own bug description was wrong (only quantities below 0.005 fail). 32 minutes, 224 steps, 427 RUB ($5.30).
- Case 2, online card payments silently not loading: the fix matched 1C's patch line for line, including the same standard attribute-check call, and the agent found the same bug in two more document types. 25 minutes, 160 steps, 155 RUB ($1.93).
- Case 3, fast-payment (SBP) button failing without a bank account: the agent blocked the operation with a clear message; 1C's patch fixed the underlying typo and kept the payment path open, which the author rates as more precise. 47 minutes, 232 steps, 329 RUB ($4.08). Score 4 of 5.
- Case 4, two features from release 3.0.14 (create a price-setting document from a supplier order; a revaluation print form), with the agent testing in the live UI and taking screenshots: the business task worked, but the agent wrote a draft record to the database and did not use the standard subsystem library mechanisms. 2 h 1 min, 420 steps, 747 RUB ($9.28). Score 3+ of 5.
- Agent permissions: read the configuration as XML and query data; load, apply, roll back and delete extensions; run external data processors.

## How they got it

Blind comparison on a demo database: official 1C patches (EF_ files) downloaded into a closed folder, the agent given only the user-facing symptom, then the agent's code compared with the patch. Scores out of 5 are the author's judgment. Costs are at DeepSeek API rates.

## Limits

- Four cases, one model, one run each; no failures or abandoned runs reported.
- Scoring is the author's own; the agent console used is not named in detail.
- Bugs picked from a published list may be simpler than unreported production bugs, and the answer exists in public patches the model could in principle have seen.

## Why it matters for companies

For companies on packaged ERP, the result suggests a fast interim fix for a known bug can cost a few dollars and under an hour, delivered as a removable extension that leaves the vendor configuration supportable. The case 3 and 4 gaps show where review is still needed: the agent solves the symptom, while the vendor fixes the cause and follows platform conventions. The blind-test setup, with a held-back official fix as ground truth, is a cheap way to evaluate any coding agent on your own stack.

## Key passages in translation

> Отдаём агенту одну строчку описания симптома из таблицы багов на ИТС. Причём описание дано как проявление ошибки у пользователя, без технических деталей.

We give the agent one line describing the symptom from the bug table on ITS. The description is how the error shows up for the user, with no technical details.

> Разумеется, никакая LLM не заменит обновления от вендора.

Of course, no LLM will replace updates from the vendor.

> Агенту не хватило знаний штатных механизмов БСП для печати и ввода на основании без записи черновика, но бизнес‑задачу он закрыл.

The agent lacked knowledge of the standard library mechanisms for printing and creating documents without writing a draft, but it did close the business task.

Original: https://habr.com/ru/articles/1102948/