Microsoft wants AI ‘agents’ to work together and retain information
Microsoft envisions AI agents collaborating across companies and improving memory retention using structured retrieval augmentation. At its Build conference, it plans to promote open standards like the Model Context Protocol to foster an interconnected “agentic web.” Microsoft’s top technologist said Sunday, ahead of the company’s annual software development conference, that the company envisions a future when its AI agents may work together with agents from other companies and remember past contacts more clearly. At its Build conference in Seattle on May 19, analysts expect Microsoft to unveil its latest tools for developers building AI systems. Microsoft is dedicated to encouraging the adoption of industry-wide standards that will allow agents from different manufacturers to collaborate, Chief Technology Officer Kevin Scott told reporters and analysts at the company’s Redmond, Washington, headquarters before to the conference. Agents are AI systems that are able to do specific activities, like fixing a software defect, on their own. Scott claims that Microsoft is backing the open-source Model Context Protocol (MCP) technology, which was first introduced by Google-backed Anthropic. According to Scott, MCP could lead to the development of an “agentic web” in a manner akin to how hypertext protocols facilitated the internet’s growth in the 1990s. “It means that your imagination gets to drive what the agentic web becomes, not just a handful of companies that happen to see some of these problems first,” Scott stated. According to Scott, Microsoft is working to improve AI agents’ recall of tasks that customers have requested of them, but as of right now, “most of what we’re building feels very transactional.” However, because it requires more computing power, enhancing an AI agent’s memory is costly. Microsoft is focusing on a new approach called structured retrieval augmentation, where an agent provides a roadmap of the subjects discussed by summarising each turn in a user discussion. “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,” Scott stated.