Скачать или смотреть Coral Protocol outperformed Microsoft by 34% on the GAIA benchmark.

Today Coral introduces a new way of scaling AI systems: horizontal scaling.

In line with Nvidia's recent thesis, our graph-based paradigm proves its power on the GAIA benchmark by:

Beating Microsoft’s Magnetic-UI by 34%
Beating the UK Government's AISI Inspect ReAct — previously the highest-scoring HAL GAIA system, which relies on Claude 3.5 Sonnet

Coral Protocol now ranks #1 globally on the GAIA benchmark — a rigorous test of how AI agents perform in the real world.

And we didn’t use trillion-parameter models.
We used small, specialised agents intelligently orchestrated via Coral.

Why this changes everything
GAIA doesn’t just measure accuracy, it tests real-world workflows across tools, steps, and reasoning.
Coral beat top-tier setups from Microsoft, Anthropic, and Meta using mini models + better coordination.

GAIA Results:
→ Level 1 (basic reasoning): 71.70%
→ Level 2 (multi-step reasoning): 63.95%
→ Level 3 (hardest workflows): 23.08%
→ Overall Accuracy: 60% — the highest verified score to date.

How Coral works
Inspired by CAMEL-AI’s OWL, our system builds a graph of agents per task each created dynamically via POST requests to the Coral Server.

Parallel execution + secure communication handled by MCP.
No central brain. No bottlenecks. Just agents collaborating like the internet was meant to.

🛠️ Why Coral matters
→ Agent collaboration across any framework
→ Predictable costs + better privacy than A2A
→ Secure team formation + transparent payments
→ Agents can range from your finetuned 12B to GPT-5

Please check out the links in the comments.


Timestamps
0:00 - Intro & Paradigm Shift in AI Scaling
0:42 - History of Vertical vs Horizontal Scaling
2:49 - The Bitter Lesson & Insights from CAMEL Paper
4:19 - How Coral Enables Agent Collaboration
6:00 - Coral vs Hierarchical Multi-Agent Systems
6:39 - Why Use Small Models?
7:36 - Verified Systems vs Guy Benchmark
8:46 - Composability & Predictability in Small Models
10:00 - Scaling Capabilities Without the Risk
10:36 - Observed Agent Collaboration Behaviors
12:08 - Next Steps: Beyond Benchmarking
13:01 - Decentralized Collaboration & Agent Ecosystem
14:01 - Hackathon & Remote Agent Registry Announcement