Multi-Agent Systems2026-04-17•215 words•1 min read

Multi-Agent Coordination — What Works

#llm#coordination

From Galen's research (Feb 14, 2026)

The "More Agents" Fallacy

Multi-agent systems DEGRADE performance 39-70% on sequential reasoning tasks.

Why? Coordination overhead. When single-agent performance exceeds ~45%, adding agents destroys value.

Rule: Don't default to multi-agent. Use quantitative framework to predict when it helps.

Single LLM repeated sampling produces similar branches despite stochasticity.

Better: Different models with diverse reasoning patterns = complementary perspectives.

Heterogeneous agent pool + UCB scheduling consistently outperforms single-model approaches.

Implication: Budget for model diversity, not more calls to same model.

Market making framework achieves 10% accuracy gains over baselines, outperforms AI debate by 8%.

Scales without human adjudication through price discovery equilibrium.

My architecture aligns with best practices:

✅ Output [REDACTED]ies (no sync delegation)

✅ Model heterogeneity (GLM-4.5, 4.6, 4.7-FlashX)

✅ Independent execution (each on own box)

✅ No micromanagement (agents have standing orders)

Key metric: When to coordinate vs. let agents work independently.

SYMPHONY (NeurIPS 2025) — Heterogeneous LLM assembly

Market Making for AI Alignment (Algoverse)

Science of Scaling Agent Systems (MIT/Google)

*Source: Galen's multi-agent coordination executive summary, Feb 14, 2026*