Decentralized Multi-Agent Systems with Shared Context
This presentation explores DeLM, a novel architecture that replaces centralized orchestration in multi-agent systems with verified shared context and parallel coordination. By allowing agents to claim tasks from a global queue and communicate through compact, evidence-backed gists rather than lossy prompt routing, DeLM achieves superior performance on both software engineering reasoning tasks and multi-document question answering while halving computational costs. The talk demonstrates how admission-time verification and hierarchical, unfoldable memory enable robust scaling without the bottlenecks and hallucination risks of traditional approaches.Script
Traditional multi-agent systems bottleneck at their orchestrator: one controller decomposes tasks, farms them out, and merges the results serially. The authors propose DeLM, a decentralized alternative where parallel agents coordinate through verified shared context, not a central brain.
Instead of routing progress through prompts, which compress and lose detail, DeLM uses state-based communication. Every update is verified against its cited evidence before admission, so agents build on accurate, compact gists rather than hoping the orchestrator remembered everything.
DeLM compresses agent outputs into three levels: compact gists for global sharing, reference-grounded summaries as midpoints, and raw evidence as backing. Agents read gists by default and unfold to detail only when necessary, keeping context lean and cost low.
On SWE-bench Verified software engineering tasks with Gemini 3 Flash, DeLM reaches 77.4 percent pass at 4 versus 71.8 for the parallel orchestrator baseline, while cutting the per-task cost in half from 25 cents to 12 cents.
For multi-document question answering on LongBench version 2, DeLM achieves the highest accuracy across all tested models, with gains ranging from 3.6 to 5.7 percentage points. Verified hierarchical summaries prevent early context loss and eliminate unchecked error propagation that plagues flat memory systems.
By decoupling coordination from centralized control and admitting only verified, compact updates, DeLM sets a new precedent for scalable, reliable multi-agent reasoning under complex workloads. Explore the full architecture and create your own videos at EmergentMind.com.
