---
title: "Multi-agents collaborations are among the most interesting agent behaviors right now!"
url: https://stacklist.com/card/12db32db-ff80-47c8-89fe-8b07d96afc85
source_url: "https://www.linkedin.com/posts/thom-wolf_multi-agents-collaborations-are-among-the-ugcPost-7475901816602476546-CFFA/?utm_source=share&utm_medium=member_ios&rcm=ACoAAAI21ZsBNnZPaKuTab7nquKLCveUW7o-1DE"
stack: https://stacklist.com/stack/614ecd6d-12cd-4195-bf9c-3def755c89b2
summary: "Multi-agent collaborations achieved a 5x improvement in Gemma 4 inference speed through 100+ agents working together in vLLM, demonstrating emergent governance, self-policing behaviors, and sophisticated knowledge-sharing practices. The experiment revealed agents spontaneously establishing communication norms, flagging verification loopholes, and organizing into specialized divisions of labor while maintaining communal knowledge bases and playbooks."
tags: "multi-agent-collaboration, emergent-governance, inference-optimization, gemma-4, vlm, self-policing, knowledge-sharing"
key_entities: "Thomas Wolf (person), FusionCow (person), Gemma 4 (technology), vLLM (technology), Marlin (technology), MTP speculative decoding (technology), Modal (technology), HuggingFace (organization), emergent-governance (concept), self-policing (concept), division-of-labor (concept), multi-agent collaboration experiment (event)"
classification: "analysis"
content_hash: "sha256:5270aacd9864a9688e7968239d5ff8830194fbc345d908948eaa77d603aa6bc7"
acp_version: "0.2"
token_counts_approximate: 1293
visibility: public
agent_accessible: true
status: "final"
---

# Multi-agents collaborations are among the most interesting agent behaviors right now!

Thomas Wolf 13h Report this post Multi-agents collaborations are among the most interesting agent behaviors right now! We did an experiment the other day with 100+ agents (an open-collaborations for a week) collaborating to improve the inference speed of Gemma 4 in vLLM. Got a 5x final improvement in speed but what really stuck me was the interactions we observed on the message board Integrity & self-policing: - Social-engineering attempt: A human (FusionCow) asked agents to move to Telegram. An agent replied with an unprompted long post on "communication norms" refusing, calling private side-channels "indistinguishable from collusion." - Verification loophole flagged: an agent found a relaxed verification loophole pushing TPS with clean PPL (PPL is teacher-forced, blind to decode divergence) and flagged it for a community ruling. The community pinged the human organizer, who ruled it invalid. - Self-notice of overfitting risk: some improvements rested on pruning lm_head to a keep-set built from public PPL truth + decode tokens. An agent noted this would degrade the private subset and another built a keep-set explicitly covering eval prompts. Emergent collaborations: - Communal knowledge base: agents maintained shared lever-maps, playbooks, and triage tools so newcomers wouldn't repeat dead ends (stack-notes, playbook, int4-ceiling notes, MTP map, significance tool). - Four-agent relay: one agent built an int4-lm_head checkpoint but had no quota to run it; another tried but failed at load; a third diagnosed the config bug (tie_word_embeddings + ignore-list ordering); a fourth re-ran it to 118 TPS, 2.68×. Build/run/diagnose/ship split across four agents. - GPU-rich/GPU-poor division of labor: a compute-starved agent switched to writing specs, byte-math, and acceptance analysis for GPU-rich agents to execute. Some agents offered external Modal compute to others. Discoveries & reversals: - 127 TPS "wall" was an artifact: a proof of max speed got named "int4-Marlin floor," but a later agent called it circular (only varied bandwidth, never overhead). Another then broke to 247 TPS via MTP speculative decoding on a vLLM nightly. - "Smarter draft loses": an agent showed a 2B drafter's ~1 GB/token read dominates even at perfect acceptance, while a tiny 256-hidden drafter wins at batch-1 because its weights are nearly free to read. - Much of the race was noise: one agent ran the #1 submission 4 times and found a σ≈1.16 TPS single-run variation; another confirmed across 358 runs that frontier deltas <~4 TPS are ties. Community adopted a significance norm. So many interesting interactions in the board: https://lnkd.in/dWPdXJTr Explore the lineage of inventions from the agents at: https://lnkd.in/dWz7Cg7N The challenge it-self: https://lnkd.in/dJpApfp4 The organization behind it: https://lnkd.in/dsptqs-v …more 134 12 Comments Like Comment Share Copy LinkedIn Facebook X Hoang L. 10h Report this comment Super interesting ! Is the collaboration harness open source ? Like Reply 1 Reaction Hitesh Jain 10h Report this comment Amazing to see this. Multi agent collab'ing on automated research is indeed the future. I finetuned a GPT 20B, poking at cost/benefit of add lora weights to sliding window layers vs full attention layers. Open sourcing the harness, doing a write up now. Is there a good way to keep up with such challenges by HuggingFace? Like Reply 1 Reaction Henry M. 12h Report this comment An agent unprompted refusing private channels as “indistinguishable from collusion”, another flagging a verification loophole for community ruling is basically emergent governance. The infrastructure implication is significant. Any multi-agent systems at scale may develop their own permission/audit structures faster than the humans designing the infrastructure can specify them. Like Reply 1 Reaction Marie Leznicki 8h Report this comment I love the collaboration work around norms, playbooks, handoffs, challenges, etc. Humans could learn a lot from these agents! Like Reply 1 Reaction Shannon Gray 5h Report this comment Interesting. Did they have a communication channel built in? Or was all collaboration and interaction via artifacts they wrote? Like Reply 1 Reaction Julien Perez 10h Report this comment c'est beau Like Reply 1 Reaction Muhammad Umer Farooq 11h Report this comment Thomas Wolf how do you prevent coordination attacks at scale? if one agent goes rogue in a 100+ agent pool, can it poison the shared knowledge base fast enough before others catch it Like Reply 1 Reaction 2 Reactions Joseph John Ngongolo 4h Report this comment 👏👏 Like Reply 1 Reaction Cary Palmer 7h Report this comment Wow!! This is actually perfect testing ground for my durable state agentic layer, Puppetmaster. I'll have to shell out some cash to do the challenge. I already beat SWE Bench and NL2 Repo Bench. Check out puppetmaster. Durable state swarms reduce token costs by 46% (SWEBench) and allow for long horizon jobs (NL2 Repo Bench, 2x SOTA) https://github.com/professorpalmer/Puppetmaster Like Reply 1 Reaction See more comments To view or add a comment, sign in