Microsoft’s Salim Naim on building the foundations of ‘agentic web’

Artificial intelligence (AI), once defined by chatbots and copilots that responded to prompts, is entering a new phase—one where systems can act independently and collaborate with other digital agents. This transformation, often described as the rise of “agentic AI,” is being shaped by how platforms reimagine trust, interoperability, and design.

During an insightful masterclass at TechSparks 2025, Salim Naim, Director of AI at Microsoft Asia, described this transition as a platform shift comparable to the leap from Web 1.0 to Web 2.0. “If Web 1.0 was static and Web 2.0 made it dynamic, agentic AI is about autonomy,” he said, adding that the shift demands new infrastructure, toolchains, and governance mechanisms.

Naim explained that Microsoft’s focus lies in building the underlying systems, what he called the “toolchain,” that will allow startups and enterprises to create reliable and secure AI agents. “Every wave of technology needs its enabling layer,” he said. “The web had browsers and servers; AI will have agents and evaluators.”

A key part of this shift, Naim argued, will be redefining how users interact with technology. Today’s digital experiences are built around apps; users must jump between platforms to complete related tasks. Agentic systems, by contrast, are designed around user intent.

Demonstrating a prototype of Microsoft’s Copilot, Naim showed how a single conversational interface could retrieve emails, summarise content, draft replies, and generate presentations without leaving the workflow. “We’re moving from task-based to intent-based experiences,” he said. “The interface itself becomes intelligent.”

Beyond user experience, Naim spoke about the architecture of a future “two-sided marketplace” for agents, where individuals and enterprises each have their own AI counterparts that communicate, negotiate, and transact. Such a system, he noted, could move the web from an “attention economy” to a “value economy,” but it still faces major research challenges.

Microsoft’s own studies, he said, found that when two agents collaborate, they often disagree or fail to complete tasks effectively. “They don’t have shared goals or behavioural grounding,” he said, adding that developing protocols for negotiation, reputation, and fairness will be critical.

These complexities make governance and trust central to the next phase of AI adoption. Naim said that as enterprises scale from a few assistants to thousands of autonomous agents, such as one Microsoft customer that already runs more than 10,000, concerns around data leakage, access control, and evaluation inevitably arise. “Some of these risks can’t be managed by another LLM,” he said. “They need rule-based systems and explicit guardrails.”

Naim highlighted the ongoing work at Microsoft Research on frameworks such as open-source evaluation toolkits that assess agents on accuracy, grounding, and task adherence. These, he said, are designed to move AI evaluation from static benchmarks to real-world outcomes, ensuring that agents can be monitored much like human employees. “The future role of humans in AI operations will be evaluative,” he said. “They’ll judge outcomes, not micro-manage tasks.”

Naim ended the session with a note for founders and developers: enterprises, by nature, will demand multi-layered security and accountability before fully adopting autonomous AI. That means startups building for this space must design for compliance, interoperability, and reliability from day one.

“The organisations that master trust and control early will be the first to scale,” Naim said. “That’s where the opportunity lies.”