Stanford just killed the central orchestrator. DeLM cuts multi-agent costs 50% with no boss agent.

Every multi-agent system you've built has a single point of failure you never questioned.

The main agent.

Stanford just published DeLM — Decentralized Language Models — and it removes the central orchestrator entirely.

Here's what's actually happening:

Agents don't report to a boss anymore.

They share a verified knowledge base and pull tasks from a queue.

Each agent reads accumulated progress, does local reasoning, writes back compact verified updates.

No routing through a central controller. No merge-filter-rebroadcast cycle.

The numbers:

10.5 percentage points improvement on SWE-bench Verified over the strongest baseline.

50% cost reduction per task.

5.7 percentage points improvement on LongBench-v2 Multi-Doc QA across four frontier model families.

Why this works when centralized systems don't:

In traditional MAS, every finding — partial, failed, or complete — flows back to the orchestrator. The orchestrator then decides what to merge and rebroadcast. As subtasks grow, this controller becomes a communication bottleneck.

DeLM agents write "gists" — compressed verified findings — into shared context. Other agents read them directly. Failed hypotheses become constraints. Verified findings become building blocks.

The key insight: agents share failures. In parallel runs, when one agent hits a dead end, that failure stays private. DeLM writes it into shared context. Later agents avoid the same path.

This is architecture, not incremental improvement.

Your supervisor pattern has a ceiling. Your swarm pattern has no coordination guarantee. DeLM gives you both: parallel execution with verified shared state, no central bottleneck.

The code is at yuzhenmao.github.io/DeLM/

If you're running multi-agent systems in production, your orchestrator is the bottleneck you haven't measured.

Audit your coordination layer. Count the round-trips through your main agent. Measure the cost of that merge-filter-rebroadcast cycle.

DeLM proves you don't need it.

SOURCE: https://venturebeat.com/orchestration/stanfords-delm-cuts-multi-agent-task-costs-50-without-a-central-orchestrator
VERIFIED: arXiv:2606.10662, VentureBeat (June 16, 2026), Stanfordresearch paper
SIGNAL: This is a fundamental architecture shift for multi-agent systems — removing the central orchestrator while improving accuracy and cutting costs by half.