Mh0tz

MeaningTrace is a semantic integrity debugging and research tool for multi-agent AI workflows, built with Python, Streamlit, and Meta Muse Spark 1.3. When AI agents collaborate, human instructions pass through multiple stages. One agent may interpret a request, another may develop a plan, and another may simulate an action. During these handoffs, important human requirements can be rephrased, shortened, or omitted. MeaningTrace makes this process visible. It records actual model inputs and outputs across three sequential roles: Interpreter, Planner, and Executor Simulator. Developers can inspect each handoff and compare the agent responses against the original human requirements. The project also investigates two experimental conditions: unprotected workflows, where downstream agents rely on the previous agent's message, and protected workflows, where they additionally receive the original human constraints. An interactive Streamlit application allows users to explore saved experiments and compare model behavior. A five-scenario pilot study includes 30 real Meta model calls. MeaningTrace is an exploratory research prototype, not a production safety system. Its goal is to make AI-agent communication more transparent and help developers investigate potential changes in meaning without assuming that every wording difference represents an error.