STANFORD JUST CUT AI AGENT COSTS IN HALF WITH A FRAMEWORK THAT LETS MACHINES COORDINATE WITHOUT A BOSS Multi-agent AI systems have a traffic problem. Every time one AI agent needs to coordinate with another, the message typically routes through a central orchestrator that acts as a dispatcher, processor and decision-maker all at once. That bottleneck drives up latency, inflates token consumption and adds cost at every layer. Stanford researchers think they have a fix. The system is called DeLM, for decentralized language model, and it is built on a simple premise: agents do not need a central controller if they share a common communication substrate. Instead of routing updates through a master agent, DeLM gives each agent access to a shared knowledge base where verified progress is logged and available for any other agent to build upon directly. The results from testing are hard to dismiss. DeLM cut multi-agent inference costs by 50 percent compared to centralized orchestration approaches and beat the SWE-bench software engineering benchmark by 10.5 percentage points. The framework also allowed agents to explore different reasoning paths in parallel without duplicating work or waiting for a controller to clear the next step. Multi-agent AI is the direction the entire industry is moving. Production deployments are scaling up. Infrastructure costs are scaling with them. If the coordination problem can be solved at this level of efficiency, the pace of deployment across enterprise software, research pipelines and autonomous systems will accelerate faster than most forecasts are currently pricing in. Keywords: Stanford DeLM, multi-agent AI, AI agent costs, AI orchestration