DeepSeek fixed 3 bugs in 1C in 104 minutes and reproduced two patches exactly

The 1C update has not been released yet, but accounting can no longer post a document. It is a familiar situation: the error is hidden somewhere in a huge standard configuration, it cannot be modified directly, and a specialist must first trace a long chain of calls. In the experiment, DeepSeek was given four such tasks and allowed to read an exported configuration, access a demo database, build extensions, and verify the results in the 1C interface.

The outcome proved more useful than another debate over whether AI will replace programmers. The agent fixed three localized defects in 25–47 minutes at a cost of RUB 155–427. In two cases, its code matched the official 1C patches exactly. But the model handled a larger enhancement only at the functional level: the solution worked, although 1C's built-in architecture was cleaner.

How the agent found the errors

In the first case, printing a universal transfer document failed with a division-by-zero error when the product quantity was below 0.005. The cause was not a zero source value. The code rounded the quantity to two decimal places immediately before division, turning 0.001–0.004 into zero. DeepSeek found the relevant query in a module of roughly 12,000 lines, added a safeguard, and built an extension. The official patch solved the problem slightly differently by retaining three decimal places. The agent also noticed that the original description of the bug's boundary was inaccurate.

In the second case, an online payment for a sales invoice silently failed to load into the database. The common handler expected the SiteAddress attribute, which was absent from some documents. The model added a check for the attribute's existence before accessing it. This solution matched the 1C patch exactly. Along the way, the agent found the same defect in two other document types, even though the original task did not mention them.

The third bug was hidden in the Faster Payments System payment button. When the bank account was empty, a typo inside a ternary operator returned a Boolean value instead of a structure, after which the code crashed when accessing a member through dot notation. DeepSeek stopped the scenario earlier and displayed a clear warning. The crash disappeared, but the official patch was more precise: it corrected the typo itself and preserved the ability to select a connection through a payment aggregator.

Where the model fell short of 1C developers

The fourth task was not a repair but a product enhancement: create a price-setting document from a supplier order and add a repricing print form. The agent built an extension, connected the commands, and independently ran the scenario in the user interface. It even added the percentage price change and color highlighting.

However, the model had a weaker grasp of the Standard Subsystems Library mechanisms. To create the document, it initially saved a draft, and it built the print form manually. The 1C developers used the standard create-based-on mechanism and standard printing with export to PDF and Excel. The feature was implemented, but the built-in version is easier to maintain.

Practical takeaways

The main conclusion is not that DeepSeek can replace a 1C developer. The experiment does not demonstrate that. It shows a narrower and more practical scenario: an agent can quickly locate a defect, build a temporary extension, and test it in a live interface while the team waits for a vendor update.

Such a fix cannot be accepted merely because it resembles the official patch. It requires a backup, a separate test database, verification of the business scenario, and the ability to disable the extension instantly. For localized incidents, the model already looks like a useful partner. For architectural enhancements, it still needs a specialist who knows the platform and can stop a solution that works but is fragile.