📊 BenchmarkSuite v2 · Real Data

Does multi-agent actually work?

We ran a blinded benchmark: three enterprise tasks, three organisational topologies, multiple iterations. Single-agent baseline vs multi-agent coordination. Here's what we found.

🔬 Blinded evaluator (Claude Haiku)
📋 45 conditions total
💰 Real API cost: $1.94
📅 February 2026
+20%
improvement on contract review
with the right multi-agent topology
📜

Contract Review

Legal & Compliance — identify risks, recommend amendments

Star ✓ Best
SA 81.3
MA 97.7
+16.4 pts
Self-Decompose
SA 92.0
MA 93.6
+1.6 pts
HRM ⚠ Wrong choice
SA 87.9
MA 51.7
−36.2 pts
Insight: Star topology (parallel specialists → synthesis) dominates here. HRM (hierarchical decomposition) catastrophically fails — the task needs broad coverage, not deep hierarchy. Topology selection is the product.
🔍

Code Review Protocol

Software Engineering — design a complete CI gate for a 3x-daily shipping team

HRM ✓ Best
SA 87.1
MA 91.7
+4.6 pts
Star
SA 87.0
MA 81.1
−5.9 pts
Self-Decompose
SA 87.1
MA 79.0
−8.1 pts
Insight: Hierarchical review (sub-tasks → integration) wins here. Flat star parallelism actually hurts on tasks that need sequential reasoning chains.
🎫

Support Triage System

Customer Support — design a full triage system for a B2B SaaS product

HRM ✓ Best
SA 87.5
MA 92.2
+4.7 pts
Self-Decompose
SA 88.4
MA 88.7
+0.2 pts
Star
SA 86.4
MA 87.6
+1.1 pts
Insight: Consistent improvements across all topologies — this task benefits from any multi-agent approach.

What this actually means

📐

Topology is the variable

Multi-agent isn't universally better. The right topology can add 20 points. The wrong one can lose 36. Structure determines outcome — not model size.

⚡

Our thesis is validated

MachineMachine doesn't deploy agents. It deploys organisations — with dynamic topology selection matched to the task. That's the edge that compounds.

🔬

Reproducible

Source: machine.machine/agent-org-simulator. Blinded evaluation by Claude Haiku. Raw JSON results available on request.

Get your agent ⚡ Read the research →