Inženýrství produkčních AI agentů: Evaluace řízení nástrojů a kontextu v marketingových procesech na základě trasování

This thesis addresses the engineering disciplines required for reliable AI agents in production environments. Model choice alone does not produce reliability in production e-commerce environments. The primary objective is to design, implement, and empirically evaluate an engineering methodology that treats agent reliability as a systems-engineering problem within a live marketing-workflow setting. The research adopts a design-and-validation strategy within the empirical domain of multi-market e-commerce marketing workflows spanning content generation, Product Discovery, and structured data transformation tasks. Agent design is decomposed into three distinct dimensions: context provisioning, tool governance, and orchestration. The methodology is instantiated in a deployed production artifact and evaluated through a trace-based harness across five architectural configurations, holding the model version constant to isolate the effect of engineering decisions on reliability. The original contribution has three linked parts: a deployed reference architecture for multi-market marketing workflows, an engineering methodology isolating three separately measurable design dimensions, and a trace-based evaluation harness enabling controlled attribution of reliability changes across the V0→V4 progression. The author designed and implemented a layered context architecture replacing monolithic system prompts, a resilience-oriented tool governance layer based on the Model Context Protocol (MCP), and a hierarchical multi-agent orchestration framework.