Master's thesis - LLM agents · In progress
Dynamic Active Toolset Management
I am researching how an LLM agent can retrieve and manage a smaller active toolset from a larger MCP-based tool pool.
01
The challenge
Large tool catalogs increase prompt cost and make correct tool selection harder. The thesis studies when and how an agent should retrieve, activate, and evaluate tools.
02
The approach
The implemented foundation uses FastMCP, the MCP protocol library, LiteLLM, and SLURM/HPC experimentation. LangGraph, FAISS, and pgvector remain planned directions rather than completed thesis components.
Keep implemented and planned work separate
The public story names the current stack precisely and labels future architecture as planned, preserving the experiment's real state.
Evaluate the system, not the demo
The research framing compares dynamic tool management with controlled baselines across success, cost, latency, retrieval, and tool use.
03
The outcome
An experiment-oriented system with explicit baselines and planned measures for task success, prompt efficiency, latency, retrieval quality, and tool-use effectiveness.