Agentic artificial intelligence infrastructure startup Coral Protocol has achieved a new record score on the popular GAIA benchmark, demonstrating that smaller AI models can outperform larger ones with the right architecture.
The open-source Coral Protocol focuses on horizontal scaling to elevate AI algorithms beyond their usual capabilities, contrasting with the prevailing industry wisdom that more parameters equate to better results. While companies like OpenAI, Google, and Microsoft Corp. develop ever-more powerful large language models, Coral believes that small language models and secure, parallel multi-agent coordination can achieve the same results.
Coral’s GAIA Agent System was developed for the GAIA benchmark, a widely recognized test for measuring agentic AI systems’ ability to solve real-world tasks. The GAIA test consists of 450 challenging questions that evaluate an AI system’s ability to act as a “general-purpose assistant” by conducting intensive research, analyzing data, and reasoning to draw conclusions.
The GAIA Agent System is based on an open-source, multi-agent collaboration framework called OWL (Optimized Workforce Learning), developed by the CAMEL-AI community. OWL automates complex tasks by coordinating dozens of specialized AI agents to work as a team. Instead of a single, monolithic LLM performing every task, Coral’s system delegates tasks to different agents, each with its own decision logic, toolkit, and specialized skills.
Coral’s system comprises numerous AI agents specialized in tasks such as planning, problem-solving, answer finding, critique, image analysis, assistance, information search, web browsing, and video analysis. These agents communicate using the Coral protocol’s Modell Context Protocol-based communication tools.
The results achieved by Coral’s system illustrate that “many heads are better than one,” as it attained a record score on the GAIA Benchmark, surpassing Microsoft’s Magnetic-UI agent’s previous-best score by 34%. Coral co-founder and Chief Technology Officer Caelum Forder said the AI industry will have to pay attention to these results, stating, “The role of small models in agentic systems has been undersold to date, but the tides are starting to turn.”
Coral’s performance validates Nvidia Corp.’s earlier hypothesis that the future of AI agents lies in small language models combined with intelligent orchestration, rather than standalone large language models. Forder added that horizontal scaling is not only possible but also more practical, as smaller AI models use significantly less power than large language models.
The Coral Protocol’s graph-based infrastructure can be applied to any kind of AI system, enabling the creation of extremely powerful AI agents based on a lightweight architecture. This means AI agents can handle more data, integrate with other systems, and generate better results without the excessive costs associated with running large language models.
