The developers updated their AI manager for marketplaces like Wildberries, Ozon, and Yandex Market, moving from a rule-based system to an agentic architecture. The initial version struggled with complex user queries, failing to process 32 out of 120 requests that fell outside predefined rules.
The new agentic model independently selects and sequences tools to fulfill requests. It incorporates memory, multi-step planning, and observation capabilities. To manage risks associated with API access, the system enforces strict constraints: a maximum of 8 steps per task, a 150-second execution limit, and a cap of 6 concurrent tool calls per step.
Data security measures include limiting table sizes to 200 rows and cell content to 120 characters. The system requires explicit user confirmation for any changes, such as price updates. Offers expire after one hour to ensure data freshness, and the agent cannot execute actions autonomously without user approval.
To improve performance, the developers implemented a knowledge library containing SEO guides and analytics documentation. They also introduced a 'judge' model that evaluates agent responses on a scale of 1 to 5 based on eight standard seller tasks. This setup allows for tracking the depth of analysis and the effectiveness of implemented constraints.
