Can AI Agents Regain Control of the Internet from Big Tech?

What does the future of the internet hold? As AI companies evolve, previously open web spaces are being overtaken by digital silos controlled by commercial AI models, sidelining enthusiasts and small businesses. In response, a coalition of grassroots researchers is determined to champion an open approach to AI.

Central to this effort is the notion of AI “agents.” These are software programs that navigate the web and interact with online platforms based on human directions, such as planning holidays and making bookings. Many perceive these agents as the next stage of evolution following services like ChatGPT, yet they face significant challenges in functionality. This is largely due to the web’s design, which favors human interaction; thus, developers are recognizing that AI agents require specialized protocols to effectively engage with online content, services, and each other.

“The objective is to establish infrastructure that facilitates communication among bots, much like software does,” explains Catherine Frick from Staffordshire University, UK.

Several competing solutions to this challenge have emerged. For instance, Anthropic, the creators of the Claude chatbot, have introduced the Model Context Protocol (MCP), which standardizes the way AI models connect to various data sources and tools. In April, Google announced its own take with the Agent2Agent (A2A) protocol, offering a distinct approach to this concept.

While these methods share similarities, they have important differences. MCP focuses on standardizing AI models’ connections to external data repositories and tools, creating a secure universal channel for two-way communication—akin to having a phone number or email for messaging. In contrast, A2A expands on this by enabling autonomous agents to discover one another, exchange information, and collaborate on tasks.

For instance, you can link your AI chatbot to the code-sharing platform GitHub via MCP, yet Google asserts that A2A could enable agents to manage job interviews, conduct calls, perform background checks—all in one streamlined process, with the agent team operating simultaneously.

However, as these protocols originate from major tech companies, concerns arise that the creators of the most successful protocol might leverage it for their own commercial gain. The MCP model necessitates oversight from a central server for connections, whereas A2A comes with its own costs, assuming that authorized agents will cooperate.

“We want to prevent an ‘Agent Internet’ from evolving into yet another ‘silo alliance,'” warns Gawee Chan, who serves as chairman of the AI Agent Protocol Group. Founded in May as part of the World Wide Web Consortium (W3C) Standards Organization, Chan emphasizes the importance of inclusivity in developing this new layer of the Internet. “If we genuinely believe that AI is a transformative technology for human society, we need an open, neutral community to guide protocol development, ensuring that its future is shared by all companies, not just a select few,” he states.

In pursuit of this goal, Chang has initiated an open-source alternative to the Big Tech Agent Protocol with the Agent Network Protocol (ANP), which predates both MCP and A2A. ANP facilitates AI agents in discovering each other and establishing identities across the web, reminiscent of the early days of the internet, when individuals created personal websites and email accounts independent of large tech intermediaries. This autonomy allows ANP-driven models to function without a central authority, enabling direct communication between distinct AI models on the same device without needing internet verification.

Flick supports the emergence of open-source, non-commercial alternatives for Agent AI. “Essentially, our aim is to restore the fundamental principle of democratization to the Internet, which is how it all began,” she remarks. Without such alternatives, she warns that tech giants could create “walled gardens” reminiscent of the challenges seen in app stores and social media platforms. “If we rely on major corporations for this, they will execute it primarily to maximize profits,” she cautions.

Google claims that its protocols are designed for universal benefit. “We will continue to enhance [A2A] to tackle real-world challenges businesses face in deploying agent frameworks. At its core, it’s structured for the future’s demands,” says Rao Sarapaneni from Google Cloud.

“We have always believed in ensuring that advancements in AI serve everyone,” adds Theo Chu, an anthropologist. “When I developed MCP, I recognized that one key strategy to avoid fragmentation and vendor lock-in—which hampers the advancement of other technologies—was to make it open-source.”

Chu asserts that MCP is integrated across major platforms, including Microsoft, OpenAI, and Google. “The success of MCP will stem from its ability to expand choices rather than restrict them,” she notes. “The collective value of the ecosystem is increased for everyone.”

The W3C Group is eager to collaborate with all stakeholders to establish technical standards industry-wide, but no specific timeline has been set. “Ultimately, our focus isn’t on the triumph or failure of any one protocol but rather on the holistic growth of the agent ecosystem.”