Simulating collaborative learning is a critical yet challenging goal in educational technology. While recent Large Language Model (LLM) advancements show promise, existing approaches often rely on static error models and rigid dialogue control and are primarily designed as student-facing training tools. To address these limitations, we present an autonomous ‘zero-player’ multi-agent simulation platform, powered by GPT-4o, designed as a computational testbed for research. Our key contributions are a data-driven, probabilistic engine for modeling a realistic spectrum of student capabilities, and a fine-grained, consensus-driven dialogue protocol that fosters emergent, bottom-up collaboration. Qualitative evaluations demonstrate that our system generates sound, expert-aligned problem solutions and, critically, produces plausible collaborative dynamics, including peer-to-peer error identification and correction. Our work establishes a high-fidelity platform for studying the mechanisms of collaborative learning and lays the groundwork for future predictive tools to help educators optimize student grouping.