Multi-agent AI systems have traditionally relied on a central orchestrator: a main agent that breaks down tasks, assigns subtasks, and collates results. But this centralized approach comes with significant costs in terms of inference dollars and coordination latency. A new framework from Stanford University, called DeLM (Decentralized Language Model), challenges that assumption by enabling agents to coordinate directly without a central boss, using a shared knowledge base as a communication substrate.
DeLM is built around three core components: parallel agents, a shared context, and a task queue. The shared context stores compact summaries, or gists, of verified findings, partial results, and documented failures. Agents write their updates directly into this shared space, which later agents can read without routing through a central controller. The task queue holds pending subtasks that agents can claim independently. This design eliminates the bottleneck of a single overloaded orchestrator that must merge, filter, and rebroadcast every piece of information.
Sponsored Protocol
According to co-developers Yuzhen Mao and Azalia Mirhoseini, in traditional centralized systems every useful finding, partial finding, and failure must be reported back to the main agent, which then decides what to merge and rebroadcast. As the number of subtasks grows, this controller becomes a communication and integration bottleneck. Moreover, the orchestrator may dilute, omit, or distort useful information, leading to lost progress. DeLM avoids these issues by allowing agents to build on prior findings, avoid repeated failures, preserve constraints, and recover detailed evidence only when needed.
Real-World Performance: Half the Cost, Higher Accuracy
Benchmark results validate the approach. On SWE-bench Verified, which tests AI models on real-world software engineering tasks, DeLM outperformed the strongest baseline by 10.5% while reducing cost per task by about 50%. On LongBench-v2 Multi-Doc QA, a long-context reasoning benchmark, DeLM achieved the highest accuracy across four model families including GPT-5.4, Claude Sonnet, Gemini Flash, and DeepSeek-V4-Pro.
Sponsored Protocol
The performance gains stem from DeLM's smart sharing mechanisms. Failed hypotheses are written into shared context, preventing later agents from wasting time on dead ends. Verified constraints become binding shared state, ensuring all agents build around them. Crucially, DeLM uses an unfoldable gist system: agents see short summaries by default but can expand them into detailed evidence when needed. This coarse-to-fine access prevents context window overload while retaining reliability.
Implications for Enterprise AI
For enterprise builders, DeLM challenges the core assumption that every multi-agent workflow requires a central controller. The results show that a decentralized model is not only theoretically cleaner but also faster, more accurate, and roughly half the cost. This could transform applications like concurrent debugging, long-document analysis, and multi-document question answering.
Sponsored Protocol
Interestingly, while some companies leverage controversies to boost AI sales, Stanford's research demonstrates that true innovation lies in rethinking the architecture itself. DeLM represents a paradigm shift from centralized orchestration to distributed cooperation. For further reading, the Wikipedia entry on multi-agent systems provides background on the field.
