UIUC proposes AgentPrimitives, a reusable latent-building-block framework for multi-agent systems
机器之心 · wechat · 2026-07-21
## What the paper proposes UIUC DREAMLab introduces **AgentPrimitives**, a framework that treats recurring multi-agent coordination patterns as reusable latent building blocks rather than task-specific agents. ## The three primitives - **ReviewPrimitive**: a reusable generate → critique → revise loop. - **VotingSelectionPrimitive**: parallel candidate exploration with selection over latent representations, instead of only final text. - **PlanningExecutionPrimitive**: separates high-level planning from low-level execution for long-horizon tasks. ## Why it matters The paper argues that most MAS suffer from a natural-language communication bottleneck: each round compresses latent state into text and then re-encodes it, which adds noise, latency, and token cost. AgentPrimitives uses **KV cache-based latent communication** plus RoPE alignment so primitives can be connected inside the model state rather than through text. ## Key results Across 8 benchmarks and 5 backbone models, the system reports: - **12.0%–16.5% average accuracy gains** over single-agent baselines - **3–4× lower token usage and latency** versus text-based MAS - only **1.3×–1.6×** overhead relative to single-agent inference - more stable performance across different backbones ## Takeaway The paper’s core claim is that multi-agent systems may need their own reusable “neural-network-like” primitives, rather than hand-built workflows for every task.
More from coding & agent
- Matt Pocock says every new codebase turns legacy within days — mattpocockuk · 2026-07-21
- Meta and Unity link AI workflows to Quest development across setup, input and validation — Vjeux · 2026-07-21
- AI Engineer World’s Fair spotlights Kids Day with 87 children learning to code — steveonjava · 2026-07-21
- Looking Glass adds persistent coding sessions that can schedule their own next turns — teleport66 · 2026-07-21
- Kimi K3 looks stronger and about 5× cheaper on a frontend dashboard task — OwariDa · 2026-07-21
- A coding-agent guardrail that checks 67 security gates before the model writes code — ZyOffsec · 2026-07-21