Stanford paper: decentralized multi-agent DeLM beats Claude Code and Codex, 2.49× faster
rohanpaul_ai · x · 2026-10-08
A Stanford paper introduces DeLM, which replaces the central coordinating agent with a shared context and task queue, eliminating wasted parallelism in multi-agent systems where agents idle waiting on peers or redo duplicate work. Agents claim tasks asynchronously, publish findings immediately, and build on each other's progress.
Results on long-horizon tasks:
- Up to 2.49× faster than vanilla Claude Code and Codex baselines
- Up to +19.2 percentage points accuracy over the same baselines
- Up to +19.9 points ProgramBench test pass rate within a 120-minute budget
- Averaged across models, 1.78× faster and 12.5 points more accurate than Codex/Claude Code on Terminal-Bench 4.0
Code, 720 collaboration trajectories, and a plugin that runs DeLM directly in Codex and Claude Code are open-sourced.
Related event: Stanford's DeLM Decentralized Multi-Agent Coding Runs 2.49x Faster(3 posts)→
More from coding & agent
- Claude Haiku 5.5 beats GPT-6 Luna at matching price, but burns ~3x the tokens — Latent Space · 2026-10-08
- ThunderSyncRL: Sync agentic RL gets up to 1.9x faster by overlapping gradients with rollouts — YouJiacheng · 2026-10-08
- Unsloth trains local decision models on 3GB VRAM, lifting accuracy from 30% to 78% — evilsocket · 2026-10-08
- Ken Goldberg seeks manufacturers applying agentic robotics to real production — ZeYanjie · 2026-10-08
- Voice-first agent interaction "activates completely different zones of the brain" — granawkins · 2026-10-08
- Paul Graham: Amazon banning agents is a rare opening for an Amazon competitor — ordax · 2026-10-08