Agentic artificial intelligence infrastructure startup Coral Protocol reckons that bigger isn’t always better when it comes to scaling AI agents. The startup has just racked up a new record score on the popular GAIA benchmark, demonstrating that it’s possible to beef up the performance of AI agents using an architecture based on “intelligent orchestration” rather than throwing more processing power behind them.
The open-source Coral Protocol is focused on the idea of horizontal scaling as a means to elevate AI algorithms beyond what they’re usually capable of, in contrast to the prevailing wisdom in the industry that more parameters equates to better results.
While the likes of OpenAI, Google and Microsoft Corp. work hard on developing ever more powerful large language models, Coral believes that it’s possible to achieve the same results using small language models and secure, parallel multi-agent coordination. And now, it has the benchmark results to back up those claims.
Coral’s GAIA Agent System was specially developed for the GAIA benchmark, which is one of the most widely-recognized tests for measuring the ability of agentic AI systems to solve real-world tasks that would take hours and hours for humans to complete. The GAIA test consists of 450 taxing questions designed to evaluate an AI system’s ability to act as a “general-purpose assistant”, Coral said, and they can only be answered by conducting intensive research, carefully analyzing data and reasoning to draw conclusions.
According to Coral, the GAIA Agent System is based on an open-source, multi-agent collaboration framework called OWL, which stands for “Optimized Workforce Learning. Developed by the CAMEL-AI community, OWL is designed to automate complex tasks by coordinating dozens of specialized AI agents so they work as a team. So instead of a single, monolithic LLM performing every task or step required to solve a problem, it instead delegates those tasks to different agents, each possessing its own decision logic and armed with its own toolkit and specialized skills.
Coral’s system is made up of numerous AI agents specialized in tasks such as planning, problem solving, answer finding, critique, image analysis, providing assistance, information search, web browsing and video analysis. They talk to each other using the Coral protocol’s Modell Context Protocol-based communication tools.
The results attained by Coral’s novel system illustrate the truth behind the old adage that ‘many heads are better than one’, for it achieved a record score on the GAIA Benchmark, surpassing the previous-best score of Microsoft’s Magnetic-UI agent by 34%.
Coral co-founder and Chief Technology Officer Caelum Forder said the AI industry is going to have no choice but to pay attention to the results attained by GAIA Agent System. “The role of small models in agentic systems has been undersold to date, but the tides are starting to turn,” he promised.
Coral’s performance validates an earlier hypothesis from Nvidia Corp., which submitted a paper in June stating that the future of AI agents lies in SLMs combined with intelligent orchestration, rather than standalone LLMs. No doubt, the likes of Google, Microsoft and OpenAI may now start to reconsider their current strategies, which are focused on training ever-more powerful LLMs as they race to develop human-like “artificial general intelligence”.
Forder added that horizontal scaling is not only possible, but also more practical, as smaller AI models use significantly less power than LLMs. “The Internet of Agents is now a working reality,” he continued. “If you’re an agent developer, just Coralise it, and if you’re an application developer, build it better for less using our infrastructure.”
Coral said that its graph-based infrastructure can be applied to any kind of AI system, and opens the door for anyone to create extremely powerful AI agents based on a lightweight architecture. It means AI agents that can handle more data, integrate with other systems and generate better results, without the excessive costs associated with running LLMs.
