In this post:
- Coral Protocol surpassed Microsoft-backed Magnetic-UI by 17% in the GAIA benchmark.
- The protocol focuses on horizontal rather than vertical scaling of intelligence, which challenges the current trends in AI model development.
- Coral protocol achieved the highest GAIA score among other verified small-model systems.
Coral Protocol, a multi-agent AI system, outperformed Microsoft’s Magnetic UI by 17% today on the GAIA benchmark. The protocol approach focused on horizontal scaling rather than vertical, which sets a new challenge to the current AI model trends. Coral achieved the highest score among other verified small-model systems.
The current benchmark score set by Coral Protocol supports Nvidia’s perspective, as shown in a paper by ArXiv, that intelligently orchestrated smaller models have the potential to outperform large-scale systems without affecting their efficiency. The Coral protocol emphasized the layered integration of focused agents worldwide. It enhanced the performance of various language models by enabling secure and parallel multi-agent coordination. The models have improved reasoning, planning, and problem-solving abilities.
Coral CTO urges agent developers to ‘Coralize’ their agents
Caelum Forder, Coral CTO, revealed that the achieved milestone proves that horizontal scaling is no longer theoretical but practical, and boasts that Coral drives it effectively. He added that the Internet of Agents is a working reality and urged agent developers to ‘Coralize’ their agents. He also called on application developers to utilize the Coral infrastructure to build effectively, saving costs.
Agent developers have focused on increasing model size to manage complex tasks and beat the competition in agentic AI development. Coral’s new approach suggested a different route where smaller orchestrated models can meet or even exceed the capabilities of large-scale systems.
GAIA benchmark evaluates advanced AI models and assesses their ability to perform real-world tasks that may require human intervention, research, and analytical skills. The benchmark has 450 complex questions designed to test AI agents in problem-solving front and general assistance use cases.
The Coral GAIA agent system used in the test is built on Coeal’s open protocol and inspired by CAMEL’s OWL. It uses specialized agents to perform tasks. The system can perform search planning, image and video analysis, critique, web browsing, and answer generation. Agents communicate via the Coral Server’s MCP communication tools.
Graph-based architecture finds a breakthrough in AI agent systems
Graph-based architecture has been highlighted following the success of the Coral protocol on the GAIA benchmark. The architecture enables developers to create robust, lightweight AI agents powered by smaller models. Systems built on graph-based architecture support stronger interconnectivity among agents and more compatibility with the broader ecosystem.
Caelum Forder, Coral CTO, added that they have proven that small models can scale beyond their previously known limits and outperform the current models. He revealed his confidence in the company’s central role as a player in the future of agentic AI.
Graph-based AI has also been utilized in different applications. MIT’s Markus Buehler developed a new method that uses graph architecture and category theory to draw correlations between unrelated domains, such as biological tissues and Beethoven’s Symphony No.9. The AI system uncovered shared patterns of complexity that suggest deep connections in how natural and artistic systems are composed.
Buehler’s model also proposed real-world innovations, such as a biomaterial inspired by abstract art, proving that graph-based AI can drive creative breakthroughs.
