The AI landscape just witnessed a shift as London’s Coral Protocol announced it has outperformed Microsoft’s Magnetic-UI by 34% on the renowned GAIA Benchmark. This leap isn’t just about scoring points—it’s a signal that the age-old race to build ever-bigger AI models may be giving way to a more nimble, horizontally scaled future.
Who is London's Coral Protocol?
Coral Protocol is a pioneering, London-based technology company developing an open and decentralized collaboration infrastructure for artificial intelligence agents. Their system empowers AI agents to communicate, coordinate, build trust, and process payments autonomously—laying the groundwork for what they call the Internet of Agents.
With a headquarters in Canary Wharf, London, Coral Protocol specializes in areas including decentralized AI, multi-agent systems, large language models, Web3 infrastructure, blockchain protocols, and decentralized identity. Their mission is to create the technical foundation for safe, scalable Artificial General Intelligence (AGI) by enabling secure and efficient collaboration among multiple autonomous agents, regardless of framework or size.
The company’s infrastructure—founded on the Model Context Protocol (MCP)—allows developers to integrate any agent or model, simplify agent onboarding, and unlock composable, privacy-focused AI for a wide array of applications. This approach positions Coral Protocol as a leading advocate for collaborative, agent-based intelligence in the next generation of AI systems.
Why Is Horizontal Scaling Outperforming Traditional Monolithic AI Models?
For years, leading research labs have poured vast resources into creating larger and larger AI models, piling on billions of parameters in the quest for intelligence. However, Coral Protocol is flipping the script. Instead of building monoliths, they’ve orchestrated swarms of specialized AI agents that work together—proving that collaboration can sometimes trump brute force.
Their multi-agent system secured the highest verified score for mini-model agents on the GAIA Benchmark, a grueling test of 450 real-world, multi-step research and reasoning tasks. Coral’s CTO, Caelum Forder, underscored the significance:
“This breakthrough marks a turning point in AI infrastructure. It’s proof that horizontal scaling isn’t just possible – it’s practical.”
GAIA’s rigorous assessment mirrors the complex demands typically shouldered by human experts. Coral’s team demonstrated that small, efficiently crafted agent collectives can rival, and even surpass, traditional giants in performance—delivering results with benefits in speed, security, and efficiency.
How Did Coral Protocol Achieve This Technical Milestone?
At the heart of Coral’s triumph is a technical marvel: the GAIA Agent System. Drawing inspiration from the CAMEL OWL framework, this system unleashes a network of specialized agents—each handling critical AI roles such as answer discovery, critique, image analysis, planning, search, and even video analysis.
The magic happens through Coral’s Model Context Protocol (MCP). This protocol enables agents to coordinate using a graph-based communication architecture, ensuring tasks are split, tackled in parallel, and woven back together with precision. This isn’t a single, bloated model trying to do everything at once; it’s a smart, orchestrated conversation between focused mini-models.
What’s truly revolutionary is that this approach doesn’t depend on model size. Any capable language model—big or small—can step into the network and perform optimally thanks to the collaborative choreography enabled by MCP.
The Industry Pulse: Shifting Toward Multi-Agent Intelligence
Coral’s success isn’t happening in a vacuum. NVIDIA, an AI giant in its own right, recently championed the notion that smaller, orchestrated models are the path forward—and Coral’s win on the GAIA Benchmark is a powerful real-world validation.
The cost benefits are hard to ignore: Coral claims that its approach can offer 10-30x lower per-token costs compared to the colossal models that dominate today’s market. This makes AI more accessible for edge devices and offers built-in data privacy, a growing concern as AI is woven into sensitive applications.
Microsoft’s Magnetic-UI—its human-centric web agent platform showcased at Build 2025—scored a respectable 42.52% on the GAIA test running in full autonomous mode. Impressive, but Coral’s system not only scored higher but did so with a dramatically smaller compute bill. This could spark a rethink among enterprise artificial intelligence (AI) architects and startups alike.
The Road Ahead: The Internet of Agents
Coral Protocol envisions something ambitious: becoming the backbone layer for the “Internet of Agents”—decentralized, privacy-first, and built for robust, safe collaborations. In this new world, AI wouldn’t be owned or monopolized by a few major labs but could be distributed everywhere, safely working together while maintaining privacy and control at the edge.
As the AI community digests the implications, one thing feels certain: we’re watching the first real proof that the future of artificial intelligence may not be a single, towering model, but many minds working hand-in-hand.
