The echo chamber concept conflates structural isolation, attitude homophily, and algorithmic filtering, contributing to contradictory findings across studies. This study decomposes the concept by independently manipulating these mechanisms in a large language model agent-based simulation (LLM-ABM) with a 3 (network structure) × 2 (recommendation algorithm) × 2 (initial attitude distribution) factorial design (500 agents, 15 rounds, N = 486,000 observations). Results show that structural isolation alone produced no significant effect on either perception bias or attitude polarization. Attitude homophily increased perception bias but did not independently shift attitudes. Algorithmic filtering increased both outcomes with larger effect sizes than network structure. Pre-existing polarization was the strongest predictor and amplified all other factors. Perception bias partially mediated their effects on attitude polarization. In sum, echo chamber effects on attitudes are conditional on the concurrence of multiple mechanisms. The study also demonstrates the methodological potential of LLM-ABM for communication research.