Stanford’s DeLM — Decentralized Multi-Agent Systems with Shared Context — tries the opposite. There is no central orchestrator. Parallel agents claim subtasks from a shared queue, read the accumulated verified progress, reason locally, and write back compact verified updates to a shared context that everyone builds on. Coordination happens through the shared substrate, not through a controller.

And on the benchmarks, it holds up. According to the paper, DeLM beats the strongest baseline by up to 10.5 points on SWE-bench Verified — at roughly half the cost per task — and improves on LongBench-v2 Multi-Doc QA across four frontier model families.

I’m not reading this as “central control is bad.” The paper doesn’t claim that, and neither do I. But two things stay with me, and both sit at the center of what I write about here.

First: coordination doesn’t always need someone in charge. The reflex — in AI and in institutions alike — is that to align many actors you need a controller. DeLM is a clean reminder that a good shared, verified context plus simple rules can carry a lot of that load, and sometimes carry it better.

Second: the win here is efficiency, not a bigger model. Same models, coordinated better, at half the cost. That’s the applied layer — where durable value is migrating, and where builders who have had to be resource-efficient hold an edge.

And it rhymes with how I think the next ecosystems should be built: not one hub commanding everything, but a network of nodes building on a shared foundation, coordinating without a bottleneck. Decentralized coordination, done well, isn’t a compromise. On this evidence, it can be the better design.

Paper: Decentralized Multi-Agent Systems with Shared Context (Stanford, via arXiv).