1C development with an AI agent

A case from practice. I was working on a multi-agent system, with my head full of that work and other thoughts. A developer from another project came over and said he was exhausted from wrestling with a DCS report: he kept trying and could not make it work. I knew he was experienced and simply worn out, so I had to help.

1) I took an ERP XML export, his report, and Unica, put them into the chat, and said: here is the problem, fix it. The agent fixed it. I gave the report to the developer, but it did not work: the data had changed, but incorrectly.

2) We took a screenshot showing what was wrong and sent it to the agent with: it does not work. We received the report and gave it to the developer. The report would not open. It turned out the agent had inserted its own comment before the query, which got in the way. We removed the line manually and checked the report: it worked as expected. On the next iteration, the agent corrected that as well.

The agent work took about 15 minutes in total. I am writing it as it happened: yes, it made mistakes, but that was the result.

The task was solved there and then, without additional struggle with infrastructure, RAG, or embeddings.

I have tried different technology stacks for agent-assisted development on the 1C platform, but Unica makes automated testing the easiest to set up. At the very least, Unica is a very strong tool among comparable solutions.

It happened on a Saturday morning. After an update to a standard Retail 3.0 file-based installation, we found a bug in DataMatrix code scanning at the cashier workstation. The background code-verification operation would hang in the verification and selection window. When scanning directly at the cashier workstation, the marking icon did not appear, and the mark was absent from the receipt, so it was not sent to Chestny ZNAK. I loaded the configuration on my computer, and GPT 5.5, using Unica — I think the version was around 0.4 at the time, though I may be mistaken; this was in early July — produced fixes for these bugs as an extension. The direct-scan marking icon was not fixed, but scanning through the verification and selection form worked correctly. That was enough overall, and the customer accepted it.

Today, with Unica+Luna, I fixed an error in a service that loads regulated documents from a medical information system: 15 minutes of work plus 15 minutes refining the prompt. Yesterday I had spent an hour and a half getting nowhere while reviewing it myself without AI. After 23 years of 1C development, the feeling is decidedly mixed.

I had spent two weeks building an integration. Before the demo, the other side changed the API specification. Out of desperation, I gave Codex — with Unica — the documentation, a key, and the sources, and asked it to help. Three hours later, I already had the new integration. I am not spreading the word within the team yet…