Alice Boccadifuoco
Design and Development of a Multi-Agent LLM System for Dynamic Interface Generation and Integration with Business APIs.
Rel. Luigi De Russis, Paolo Garza. Politecnico di Torino, Corso di laurea magistrale in Ingegneria Informatica (Computer Engineering), 2026
|
Preview |
PDF (Tesi_di_laurea)
- Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (3MB) | Preview |
Abstract
This thesis investigates the engineering challenges of deploying multi-agent LLM systems in enterprise environments, where the shared context window becomes a critical bottleneck. Through an industrial case study, the Joule multi-agent integration for the Unified Analytics Overview Capability of SAP Cloud ALM, four design strategies are applied: an API redesign achieving 73–85% token reduction on the ProviderDefinitions endpoint by avoiding the Execution Trap, empirically validated against a quantitative baseline benchmark; a hierarchical Router-Skill orchestration pattern enforcing context isolation and a migration of domain knowledge to a structured MCP server implementing progressive disclosure, both evaluated qualitatively against the baseline architecture; and a proposed dual-LLM compression architecture (Gemini 2.5 Flash for context summarisation, GPT-4.1 for domain reasoning), including an isolated evaluation of the toolkit compression, which recorded a 61.5% size reduction for the target test case.
Within the scope of this case study, the results indicate that principled context management and clean architectural boundaries function as prerequisites, not optimisations, for reliable multi-agent systems in production enterprise environments
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Corso di laurea
Classe di laurea
Ente in cotutela
Aziende collaboratrici
URI
![]() |
Modifica (riservato agli operatori) |
