Ilyas Salikhov (CTO of RetailCRM) secured top positions in the Speed and Live PROD categories of the BitGN E-commerce AI Agent (ECOM1) benchmark by applying the innovative "Exoskeleton" architecture. Instead of relying on ultra-powerful and expensive models, the solution demonstrates the superiority of a hybrid approach that combines lightweight LLMs with deterministic software code.
What Happened
As part of the BitGN challenge to create efficient AI agents for e-commerce, the "Exoskeleton" architecture showed the best results in speed and real-time performance. The agent is based on the use of gpt-5.4-mini, which is 6 times cheaper than using gpt-5.5, and utilizes ultra-lightweight gpt-5.4-nano models for tasks such as classification and formatting.
Context
The traditional approach to AI agent development often focuses on trying to solve all tasks using a single powerful "super-model." However, the "Exoskeleton" architecture offers a different paradigm: the model acts merely as a "conductor" or dispatcher, while critical operations—calculations, catalog management, security enforcement, and maintaining an evidence ledger—are performed by reliable software code.
Why It Matters for the Industry
The success of this approach sets a new efficiency standard for commercial AI agents. Moving from an "LLM-only" strategy to hybrid systems (LLM + Deterministic Code) allows for a radical reduction in TCO (Total Cost of Ownership) and increased system predictability. This shifts the industry focus from the endless scaling of model weights toward system architecture and the orchestration of specialized tools.
Why It Matters for Users
For developers and businesses, this means the ability to create high-performance agents without massive API costs. Using budget models in conjunction with programmatic logic allows for achieving the required quality and speed while minimizing the risk of hallucinations in critical nodes, such as financial calculations or order management.
Sources
Author
Look at AI, Editorial Team