The developers replaced their custom AI agent core for marketplace management with Pydantic AI after finding the original version too restrictive. The previous iteration relied on hardcoded limits, such as a maximum of eight steps per task and truncated data, which prevented the model from performing complex analysis. The team deleted 582 lines of code and added 121 lines to implement the new framework.
The new architecture grants the agent full access to the Wildberries API, covering 307 distinct methods, alongside an official documentation tool. Instead of using pre-defined wrappers, the agent now decides how to utilize these tools independently. Data truncation has been removed, allowing the model to process full datasets, and the agent can now perform web searches to retrieve real-time information.
The core principle of human oversight remains: any changes to a marketplace account require explicit confirmation from the owner. However, the process has been streamlined; the agent now aggregates multiple requested actions into a single package for review. The system also supports parallel data reading and provides real-time progress updates to the user.
The transition has eliminated arbitrary constraints on the number of tool calls and steps. The agent now caches successful API calls as recipes to improve future performance and includes error recovery mechanisms that allow tasks to resume after interruptions. Initial tests show the agent is capable of deep, multi-step research tasks.
