Researchers at Carnegie Mellon University developed Message Passing Language Models (MPLMs), which enable parallel LLM threads to communicate directly with each other rather than through a single coordinator. This architecture solved structured puzzles like Sudoku and 3-SAT faster and more efficiently than earlier parallelization methods by eliminating bottlenecks at a central coordination point.
Multi-agent is a beta feature that enables a root AI agent to spawn and coordinate multiple subagents in parallel to tackle independent portions of complex tasks like codebase exploration and code review. This approach provides faster execution through parallelization, maintains focused context for each subagent, and allows model-directed orchestration without requiring application-level coordination logic.