LLM Multi-Agent Coordination
Research on coordinating LLM-based agents towards a shared goal, investigating memory architectures, inter-agent communication protocols, and explainability mechanisms in tool-rich environments.
Novel approach to agent memory and communication
Explainability in multi-agent LLM systems
Tool-rich environment orchestration
Academic research with real-world applicability
My bachelor's degree thesis at FIB-UPC investigates one of the most pressing challenges in modern AI systems: how to coordinate multiple LLM-based agents to work together towards a shared goal while maintaining transparency and explainability.
The research focuses on three interconnected pillars. First, memory architectures — how agents store, retrieve, and share information across interactions, including both short-term conversational context and long-term knowledge that persists across tasks. Second, inter-agent communication — designing protocols that allow agents to delegate tasks, share findings, and resolve conflicts without a rigid central controller. Third, explainability — ensuring that multi-agent decisions can be traced, understood, and audited by humans, which is critical for deployment in business and safety-sensitive contexts.
The work is grounded in practical tool-rich environments where agents must not only reason but also take actions — calling APIs, querying databases, manipulating files, and coordinating sequential workflows.
This reflects real-world scenarios like the AI platforms I've built professionally, where agent systems need to handle complex document processing, data extraction, and multi-step business logic.
The thesis contributes both a theoretical framework for thinking about LLM agent coordination and practical experiments demonstrating how different memory and communication strategies affect task completion, accuracy, and explainability in multi-agent setups.