The author analyzed the use of OpenClaw agents in internal workflows, specifically for routing customer inquiries and exporting data from 1C. The study highlights cases where LLM-based agents were replaced by standard Python code to ensure process reliability and prevent duplicate actions.
In the lead routing process, a nested agent caused timeouts, leading to inconsistent states where inquiries were marked as processed before messages were sent. After switching to standard Python logic and updating the state only after successful delivery, the system processed 123 inquiries without errors.
For 1C task exports, the author implemented an agent that triggers via Telegram commands or regex-based message monitoring. The system retrieves task details, comments, and attachments via MCP, then compiles a DOCX file. To date, the system has generated 53 task packages, including 17 attachments.
The author emphasizes that while agents serve as effective interfaces, critical operations requiring predictable outcomes—such as file handling, state management, and avoiding duplicate processing—are better handled by deterministic code rather than LLMs.
