DeLM Cuts Multi‑Agent AI Costs by 50%—No Central Orchestrator Needed

DeLM slashes multi‑agent costs by ↓ 50%

DeLM is Stanford’s new decentralized language model that lets autonomous agents share progress without a central controller. By storing verified “gists” in a common knowledge base, agents can read each other’s findings directly, cutting redundant work and inference spend.

Why the old orchestrator model falters

Conventional systems route every sub‑task back to a master agent, creating a bottleneck that inflates latency and dollars. As the task pool expands, the controller becomes a choke point, often distorting or dropping useful signals.

“Agents write compact, verified updates into a shared context that later agents can read directly,” Yuzhen Mao said.

DeLM replaces the hierarchy with three pillars: a shared context of concise summaries, a dynamic task queue, and parallel execution. Agents claim pending tasks, compress results into “gists,” and verify them against evidence before broadcasting.

Real‑world performance

On the SWE‑bench Verified benchmark, DeLM outperformed the strongest baseline by ↑ 10.5% and halved the cost per task. The same architecture topped LongBench‑v2 Multi‑Doc QA across four model families, including GPT‑5.4 and Claude Sonnet.

Failures are also shared. When one agent hits a dead end, its error is logged, preventing peers from repeating the mistake—a saving that compounds as the queue grows.

For enterprises eyeing scalable AI, the takeaway is clear: a decentralized approach can be faster, cheaper, and more reliable. As AI workloads expand, the model mirrors lessons learned during the pandemic about distributed resilience.

For broader industry context see Reuters and Bloomberg.

Analysis by: Dr. Aris Thorne

Artificial Intelligence Researcher