At its Build 2025 developer conference on May 19, Microsoft outlined a future where AI “agents” from different companies seamlessly interoperate and retain richer context from past interactions. Speaking from Redmond, CTO Kevin Scott emphasized that breaking down siloed, transactional AI experiences is key to unlocking more powerful, web-wide intelligence.
To foster cross-vendor collaboration, Microsoft is backing the Model Context Protocol (MCP)—an open-source initiative led by Anthropic and supported by Google. MCP aims to become to AI agents what HTTP and HTML were to the early internet: a unifying framework that lets any compliant agent discover, query, and invoke others’ capabilities. “Your imagination gets to drive what the agentic web becomes,” Scott noted, “not just a handful of companies that happen to see these problems first.”
Beyond collaboration, Scott stressed that AI agents need better “long-term” memory. Current implementations are largely transactional—agents forget past requests once a session ends. To solve this, Microsoft is pioneering structured retrieval augmentation (SRA). In SRA, agents automatically extract and index concise summaries of each conversational turn. This index then acts as a roadmap, quickly guiding the agent back to relevant user context without reprocessing entire histories from scratch. “This mirrors how biological brains learn,” Scott explained, “rather than brute-forcing every detail each time.”
With these two pillars—interoperability through MCP and memory via SRA—Microsoft hopes to accelerate the creation of more autonomous, context-aware applications. Developers at Build 2025 can expect new SDKs, code samples, and preview releases that integrate these standards, paving the way for an “agentic web” where task-specific bots cooperate and recall prior work intuitively.
