Microsoft pushes for an open “agentic web” to democratize AI collaboration

Microsoft is advancing AI interoperability by developing protocols that will allow different companies’ AI agents to collaborate effectively. The initiative focuses on creating industry standards for AI agent communication and improving AI memory systems to enhance user experiences. This push for an “agentic web” parallels how internet protocols transformed connectivity in the 1990s, potentially democratizing the future development of AI agent ecosystems beyond just a few large companies.

The big picture: Microsoft’s Chief Technology Officer Kevin Scott unveiled the company’s vision for interconnected AI agents ahead of its annual Build developer conference in Seattle.

• Scott emphasized Microsoft’s support for the Model Context Protocol (MCP), an open-source protocol created by Anthropic, as a foundation for creating an “agentic web.”
• The company believes this interoperability standard could democratize AI development, allowing developers’ “imagination to drive what the agentic web becomes, not just a handful of companies.”

Why this matters: The ability for AI agents from different companies to work together would significantly expand their capabilities and usefulness across organizational boundaries.

• Agents—AI systems that can autonomously complete specific tasks like fixing software bugs—currently operate in isolated environments, limiting their collaborative potential.
• Creating industry standards for agent communication could mirror how hypertext protocols helped spread internet adoption in the 1990s.

Key technical focus: Microsoft is developing more efficient memory systems for AI agents to improve their ability to recall previous interactions with users.

• Current AI interactions feel “very transactional” according to Scott, with agents often unable to remember context from earlier conversations.
• The challenge is that better memory requires more computing power, which significantly increases costs.

The solution: Microsoft is implementing “structured retrieval augmentation” to make AI memory more efficient and affordable.

• This approach extracts brief summaries from each conversation turn, creating a conceptual roadmap of the discussion rather than storing everything.
• Scott compared this to human cognition, noting that “This is a core part of how you train a biological brain—you don’t brute force everything in your head every time you need to solve a particular problem.”